Asymmetric default label noise

Description

Introduction of Asymmetric default label noise into a classification dataset.

Usage

## Default S3 method:
asy_def_ln(x, y, level, def = 1, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
asy_def_ln(formula, data, ...)

Arguments

Details

Asymmetric default label noise randomly selects (level[i]·100)% of the samples of each class C[i] in the dataset -the order of the class labels is determined by order. Then, the labels of these samples are replaced by a fixed label (C[def]) within the set of class labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

R. C. Prati, J. Luengo, and F. Herrera. Emerging topics and challenges of learning from noisy data in nonstandard classification: a survey beyond binary class noise. Knowledge and Information Systems, 60(1):63–97, 2019. doi:10.1007/s10115-018-1244-4.

See Also

sym_nean_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- asy_def_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                     level = c(0.1, 0.2, 0.3), order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- asy_def_ln(formula = Species ~ ., data = iris2D, 
                     level = c(0.1, 0.2, 0.3), order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Asymmetric interval-based attribute noise

Description

Introduction of Asymmetric interval-based attribute noise into a classification dataset.

Usage

## Default S3 method:
asy_int_an(x, y, level, nbins = 10, sortid = TRUE, ...)

## S3 method for class 'formula'
asy_int_an(formula, data, ...)

Arguments

Details

Asymmetric interval-based attribute noise corrupts (level[i]·100)% of the values for each attribute A[i] in the dataset. In order to corrupt an attribute A[i], (level[i]·100)% of the samples in the dataset are chosen. To corrupt a value in numeric attributes, the attribute is split into equal-frequency intervals, one of its closest intervals is picked out and a random valuen within the interval is chosen as noisy. For nominal attributes, a random value within the domain is selected.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

M. V. Mannino, Y. Yang, and Y. Ryu. Classification algorithm sensitivity to training data with non representative attribute noise. Decision Support Systems, 46(3):743-751, 2009. doi:10.1016/j.dss.2008.11.021.

See Also

asy_uni_an, symd_gimg_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- asy_int_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                         level = c(0.1, 0.2))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- asy_int_an(formula = Species ~ ., data = iris2D,
                         level = c(0.1, 0.2))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Asymmetric sparse label noise

Description

Introduction of Asymmetric sparse label noise into a classification dataset.

Usage

## Default S3 method:
asy_spa_ln(x, y, levelO, levelE, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
asy_spa_ln(formula, data, ...)

Arguments

Details

Asymmetric sparse label noise randomly selects (levelO·100)% of the samples in each odd class and (levelE·100)% of the samples in each even class -the order of the class labels is determined by order. Then, each odd class is flipped to the next class, whereas each even class is flipped to the previous class. If the dataset has an odd number of classes, the last class is not corrupted.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

J. Wei and Y. Liu. When optimizing f-divergence is robust with label noise. In Proc. 9th International Conference on Learning Representations, pages 1-11, 2021. url:https://openreview.net/forum?id=WesiCoRVQ15.

See Also

mind_bdir_ln, fra_bdir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- asy_spa_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                       levelO = 0.1, levelE = 0.3, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- asy_spa_ln(formula = Species ~ ., data = iris2D, 
                        levelO = 0.1, levelE = 0.3, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Asymmetric uniform attribute noise

Description

Introduction of Asymmetric uniform attribute noise into a classification dataset.

Usage

## Default S3 method:
asy_uni_an(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
asy_uni_an(formula, data, ...)

Arguments

Details

Asymmetric uniform attribute noise corrupts (level[i]·100)% of the values for each attribute A[i] in the dataset. In order to corrupt an attribute A[i], (level[i]·100)% of the samples in the dataset are chosen. Then, their values for A[i] are replaced by random different ones between the minimum and maximum of the domain of the attribute following a uniform distribution (for numerical attributes) or choosing a random value (for nominal attributes).

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

A. Petety, S. Tripathi, and N. Hemachandra. Attribute noise robust binary classification. In Proc. 34th AAAI Conference on Artificial Intelligence, pages 13897-13898, 2020.

See Also

symd_gimg_an, unc_vgau_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- asy_uni_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                         level = c(0.1, 0.2))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- asy_uni_an(formula = Species ~ ., data = iris2D,
                         level = c(0.1, 0.2))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Asymmetric uniform label noise

Description

Introduction of Asymmetric uniform label noise into a classification dataset.

Usage

## Default S3 method:
asy_uni_ln(x, y, level, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
asy_uni_ln(formula, data, ...)

Arguments

Details

Asymmetric uniform label noise randomly selects (level[i]·100)% of the samples of each class C[i] in the dataset -the order of the class labels is determined by order. Finally, the labels of these samples are randomly replaced by other different ones within the set of class labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

Z. Zhao, L. Chu, D. Tao, and J. Pei. Classification with label noise: a Markov chain sampling framework. Data Mining and Knowledge Discovery, 33(5):1468-1504, 2019. doi:10.1007/s10618-018-0592-8.

See Also

maj_udir_ln, asy_def_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- asy_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                    level = c(0.1, 0.2, 0.3), order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- asy_uni_ln(formula = Species ~ ., data = iris2D, 
                     level = c(0.1, 0.2, 0.3), order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Attribute-mean uniform label noise

Description

Introduction of Attribute-mean uniform label noise into a classification dataset.

Usage

## Default S3 method:
attm_uni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
attm_uni_ln(formula, data, ...)

Arguments

Details

For each sample, its distance to the mean of each attribute is computed. Then, (level·100)% of the samples in the dataset are randomly selected to be mislabeled, more likely choosing samples whose features are generally close to the mean. The labels of these samples are randomly replaced by other different ones within the set of class labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References

References

B. Nicholson, V. S. Sheng, and J. Zhang. Label noise correction and application in crowdsourcing. Expert Systems with Applications, 66:149-162, 2016. doi:10.1016/j.eswa.2016.09.003.

See Also

qua_uni_ln, exps_cuni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- attm_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- attm_uni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Distance to SVM decision boundary

Description

Calculation of the distance of each sample to the SVM decision boundary in a classification problem.

Usage

bord_dist(x, y, krn = "linear")

Arguments

Value

A vector of length nrow(x) with the distance of each sample to the decision boundary.

Mislabeling based on k-nearest neighbors

Description

Computation of a noisy label based on majority class among k nearest neighbors with different label.

Usage

bord_noise(x, y, num_noise, idx_noise, k)

Arguments

Value

A vector of length length(y) with the class of each sample, including the new noisy classes for the samples with indices idx_noise.

Boundary/dependent Gaussian attribute noise

Description

Introduction of Boundary/dependent Gaussian attribute noise into a classification dataset.

Usage

## Default S3 method:
boud_gau_an(x, y, level, k = 0.2, sortid = TRUE, ...)

## S3 method for class 'formula'
boud_gau_an(formula, data, ...)

Arguments

Details

Boundary/dependent Gaussian attribute noise corrupts (level·100)% samples among the ((level+0.1)·100)% of samples closest to the decision boundary. Their attribute values are corrupted by adding a random number that follows a Gaussian distribution of mean = 0 and standard deviation = (max-min)·k, being max and min the limits of the attribute domain. For nominal attributes, a random value is chosen.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

J. Bi and T. Zhang. Support vector classification with input data uncertainty. In Advances in Neural Information Processing Systems, volume 17, pages 161-168, 2004. url:https://proceedings.neurips.cc/paper/2004/hash/22b1f2e0983160db6f7bb9f62f4dbb39-Abstract.html.

See Also

imp_int_an, asy_int_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- boud_gau_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- boud_gau_an(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Clustering-based voting label noise

Description

Introduction of Clustering-based voting label noise into a classification dataset.

Usage

## Default S3 method:
clu_vot_ln(x, y, k = nlevels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
clu_vot_ln(formula, data, ...)

Arguments

Details

Clustering-based voting label noise divides the dataset into k clusters. Then, the labels of each cluster are relabeled with the majority class among its samples.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, which considers k-means as unsupervised clustering method.

References

Q. Wang, B. Han, T. Liu, G. Niu, J. Yang, and C. Gong. Tackling instance-dependent label noise via a universal probabilistic model. In Proc. 35th AAAI Conference on Artificial Intelligence, pages 10183-10191, 2021. url:https://ojs.aaai.org/index.php/AAAI/article/view/17221.

See Also

sco_con_ln, mis_pre_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- clu_vot_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)])

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- clu_vot_ln(formula = Species ~ ., data = iris2D)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

diris2D dataset

Description

Discretized version of the iris2D dataset.

Usage

data(diris2D)

Format

A data.frame with 103 samples (rows) and 3 variables (columns) named Petal.Length, Petal.Width and Species.

Source

Data collected by E. Anderson (1935).

References

R. A. Fisher. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7:179-188, 1936.

E. Anderson. The irises of the Gaspe Peninsula. Bulletin of the American Iris Society, 59:2-5, 1935.

See Also

iris2D, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(diris2D)

# noise introduction
set.seed(9)
outdef <- sym_uni_ln(x = diris2D[,-ncol(diris2D)], y = diris2D[,ncol(diris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)

Exponential borderline label noise

Description

Introduction of Exponential borderline label noise into a classification dataset.

Usage

## Default S3 method:
exp_bor_ln(x, y, level, rate = 1, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
exp_bor_ln(formula, data, ...)

Arguments

Details

Exponential borderline label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, an exponential distribution with parameter rate is used to compute the value for the probability density function associated to each distance. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of the probability density function. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering SVM with linear kernel as classifier, a mislabeling process using the neighborhood of noisy samples and a noise level to control the number of errors in the data.

References

J. Bootkrajang. A generalised label noise model for classification in the presence of annotation errors. Neurocomputing, 192:61–71, 2016. doi:10.1016/j.neucom.2015.12.106.

See Also

pmd_con_ln, clu_vot_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- exp_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- exp_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Exponential/smudge completely-uniform label noise

Description

Introduction of Exponential/smudge completely-uniform label noise into a classification dataset.

Usage

## Default S3 method:
exps_cuni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
exps_cuni_ln(formula, data, ...)

Arguments

Details

Exponential/smudge completely-uniform label noise includes an additional attribute (smudge) in the dataset with random values in [0,1]. This attribute is used to compute the mislabeling probability for each sample based on an exponential function (in which level is used as lambda). It selects samples in the dataset based on these probabilities. Finally, the labels of these samples are randomly replaced by others within the set of class labels (this model can choose the original label of a sample as noisy).

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

B. Denham, R. Pears, and M. A. Naeem. Null-labelling: A generic approach for learning in the presence of class noise. In Proc. 20th IEEE International Conference on Data Mining, pages 990–995, 2020. doi:10.1109/ICDM50108.2020.00114.

See Also

opes_idu_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- exps_cuni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.8)

# show results
summary(outdef, showid = TRUE)
plot(outdef, pca = TRUE)

# usage of the method for class formula
set.seed(9)
outfrm <- exps_cuni_ln(formula = Species ~ ., data = iris2D, level = 0.8)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Find the differences between two datasets

Description

Detect the differences between two datasets, focusing on the input attributes (x, xnoise), the output class (y, ynoise) or both depending on the type of the model (label, attributes, combined).

Usage

findnoise(x, y, xnoise, ynoise, model)

Arguments

Value

A list with four elements:

Fraud bidirectional label noise

Description

Introduction of Fraud bidirectional label noise into a classification dataset.

Usage

## Default S3 method:
fra_bdir_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
fra_bdir_ln(formula, data, ...)

Arguments

Details

Fraud bidirectional label noise randomly selects (level·100)% of the samples from the minority class in the dataset and level·10 samples from the majority class. Then, minority class samples are mislabeled as belonging to the majority class and majority class samples are mislabeled as belonging to the minority class. In case of ties determining minority and majority classes, a random class is chosen among them.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

Z. Salekshahrezaee, J. L. Leevy, and T. M. Khoshgoftaar. A reconstruction error-based framework for label noise detection. Journal of Big Data, 8(1):1-16, 2021. doi:10.1186/s40537-021-00447-5.

See Also

irs_bdir_ln, pai_bdir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- fra_bdir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- fra_bdir_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Gamma borderline label noise

Description

Introduction of Gamma borderline label noise into a classification dataset.

Usage

## Default S3 method:
gam_bor_ln(x, y, level, shape = 1, rate = 0.5, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
gam_bor_ln(formula, data, ...)

Arguments

Details

Gamma borderline label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, a gamma distribution with parameters (shape, rate) is used to compute the value for the probability density function associated to each distance. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of the probability density function. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering SVM with linear kernel as classifier, a mislabeling process using the neighborhood of noisy samples and a noise level to control the number of errors in the data.

References

J. Bootkrajang. A generalised label noise model for classification. In Proc. 23rd European Symposium on Artificial Neural Networks, pages 349-354, 2015. url:https://dblp.org/rec/conf/esann/Bootkrajang15.html?view=bibtex.

See Also

exp_bor_ln, pmd_con_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- gam_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- gam_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Gaussian borderline label noise

Description

Introduction of Gaussian borderline label noise into a classification dataset.

Usage

## Default S3 method:
gau_bor_ln(x, y, level, mean = 0, sd = 1, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
gau_bor_ln(formula, data, ...)

Arguments

Details

Gaussian borderline label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, a Gaussian distribution with parameters (mean, sd) is used to compute the value for the probability density function associated to each distance. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of the probability density function. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data, considering SVM with linear kernel as classifier, a mislabeling process using the neighborhood of noisy samples and a noise level to control the number of errors in the data.

References

J. Bootkrajang and J. Chaijaruwanich. Towards instance-dependent label noise-tolerant classification: a probabilistic approach. Pattern Analysis and Applications, 23(1):95-111, 2020. doi:10.1007/s10044-018-0750-z.

See Also

sigb_uni_ln, larm_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- gau_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- gau_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Gaussian-mixture borderline label noise

Description

Introduction of Gaussian-mixture borderline label noise into a classification dataset.

Usage

## Default S3 method:
gaum_bor_ln(
  x,
  y,
  level,
  mean = c(0, 2),
  sd = c(sqrt(0.5), sqrt(0.5)),
  w = c(0.5, 0.5),
  k = 1,
  sortid = TRUE,
  ...
)

## S3 method for class 'formula'
gaum_bor_ln(formula, data, ...)

Arguments

Details

Gaussian-mixture borderline label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, a Gaussian mixture distribution with parameters (mean, sd) and weights w is used to compute the value for the probability density function associated to each distance. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of the probability density function. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering SVM with linear kernel as classifier, a mislabeling process using the neighborhood of noisy samples and a noise level to control the number of errors in the data.

References

J. Bootkrajang and J. Chaijaruwanich. Towards instance-dependent label noise-tolerant classification: a probabilistic approach. Pattern Analysis and Applications, 23(1):95-111, 2020. doi:10.1007/s10044-018-0750-z.

See Also

gau_bor_ln, sigb_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- gaum_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- gaum_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Gaussian-level uniform label noise

Description

Introduction of Gaussian-level uniform label noise into a classification dataset.

Usage

## Default S3 method:
glev_uni_ln(x, y, level, sd = 0.01, sortid = TRUE, ...)

## S3 method for class 'formula'
glev_uni_ln(formula, data, ...)

Arguments

Details

For each sample, Gaussian-level uniform label noise assigns a random probability following a Gaussian distribution of mean = level and standard deviation sd. Noisy samples are chosen according to these probabilities. The labels of these samples are randomly replaced by other different ones within the set of class labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

D. Liu, G. Yang, J. Wu, J. Zhao, and F. Lv. Robust binary loss for multi-category classification with label noise. In Proc. 2021 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1700-1704, 2021. doi:10.1109/ICASSP39728.2021.9414493.

See Also

sym_hienc_ln, sym_nexc_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- glev_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- glev_uni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Hubness-proportional uniform label noise

Description

Introduction of Hubness-proportional uniform label noise into a classification dataset.

Usage

## Default S3 method:
hubp_uni_ln(x, y, level, k = 3, sortid = TRUE, ...)

## S3 method for class 'formula'
hubp_uni_ln(formula, data, ...)

Arguments

Details

Hubness-proportional uniform label noise is based on the presence of hubs in the dataset. It selects (level·100)% of the samples in the dataset using a discrete probability distribution based on the concept of hubness, which is computed using the nearest neighbors of each sample. Then, the class labels of these samples are randomly replaced by different ones from the c classes.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

N. Tomasev and K. Buza. Hubness-aware kNN classification of high-dimensional data in presence of label noise. Neurocomputing, 160:157-172, 2015. doi:10.1016/j.neucom.2014.10.084.

See Also

smu_cuni_ln, oned_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- hubp_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- hubp_uni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Importance interval-based attribute noise

Description

Introduction of Importance interval-based attribute noise into a classification dataset.

Usage

## Default S3 method:
imp_int_an(x, y, level, nbins = 10, ascending = TRUE, sortid = TRUE, ...)

## S3 method for class 'formula'
imp_int_an(formula, data, ...)

Arguments

Details

The values in level are ordered and assigned to attributes according to their information gain (using the ordering given by ascending). Then, Importance interval-based attribute noise corrupts (level[i]·100)% of the values for each attribute A[i] in the dataset. In order to corrupt each attribute A[i], (level[i]·100)% of the samples in the dataset are chosen. To corrupt a value in numeric attributes, the attribute is split into equal-frequency intervals, one of its closest intervals is picked out and a random value within the interval is chosen as noisy. For nominal attributes, a random value within the domain is chosen.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

M. V. Mannino, Y. Yang, and Y. Ryu. Classification algorithm sensitivity to training data with non representative attribute noise. Decision Support Systems, 46(3):743-751, 2009. doi:10.1016/j.dss.2008.11.021.

See Also

asy_int_an, asy_uni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- imp_int_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                        level = c(0.1, 0.2))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- imp_int_an(formula = Species ~ ., data = iris2D, 
                        level = c(0.1, 0.2))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

iris2D dataset

Description

A 2-dimensional version of the well-known iris dataset. It maintains the attributes Petal.Length and Petal.Width, which give the measurements in centimeters of the petal length and width of iris flowers belonging to three different species (setosa, versicolor and virginica). Duplicate and contradictory samples are removed from the dataset, resulting in a total of 103 samples.

Usage

data(iris2D)

Format

A data.frame with 103 samples (rows) and 3 variables (columns) named Petal.Length, Petal.Width and Species.

Source

Data collected by E. Anderson (1935).

References

R. A. Fisher. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7:179-188, 1936.

E. Anderson. The irises of the Gaspe Peninsula. Bulletin of the American Iris Society, 59:2-5, 1935.

See Also

sym_uni_ln, sym_uni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

library(ggplot2)
library(RColorBrewer)

data(iris2D)

ggplot(data = iris2D, aes(x = iris2D[,1], y = iris2D[,2], color = iris2D[,3])) +
   geom_point(stroke = 0.5) +
   xlim(min(iris2D[,1]), max(iris2D[,1])) +
   ylim(min(iris2D[,2]), max(iris2D[,2])) +
   xlab(names(iris2D)[1]) + 
   ylab(names(iris2D)[2]) +
   labs(color='Species') +
   scale_color_manual(values = brewer.pal(3, "Dark2")) +
   theme(panel.border = element_rect(colour = "black", fill=NA),
         aspect.ratio = 1,
         axis.text = element_text(colour = 1, size = 12),
         legend.background = element_blank(),
         legend.box.background = element_rect(colour = "black"))

IR-stable bidirectional label noise

Description

Introduction of IR-stable bidirectional label noise into a classification dataset.

Usage

## Default S3 method:
irs_bdir_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
irs_bdir_ln(formula, data, ...)

Arguments

Details

IR-stable bidirectional label noise randomly selects (level·100)% of the samples from the minority class in the dataset and the same amount of samples from the majority class. Then, minority class samples are mislabeled as belonging to the majority class and majority class samples are mislabeled as belonging to the minority class. In case of ties determining minority and majority classes, a random class is chosen among them.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

B. Chen, S. Xia, Z. Chen, B. Wang, and G. Wang. RSMOTE: A self-adaptive robust SMOTE for imbalanced problems with label noise. Information Sciences, 553:397-428, 2021. doi:10.1016/j.ins.2020.10.013.

See Also

pai_bdir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- irs_bdir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- irs_bdir_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Laplace borderline label noise

Description

Introduction of Laplace borderline label noise into a classification dataset.

Usage

## Default S3 method:
lap_bor_ln(x, y, level, mu = 0, b = 1, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
lap_bor_ln(formula, data, ...)

Arguments

Details

Laplace borderline label noise uses uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, a Laplace distribution with parameters (mu, b) is used to compute the value for the probability density function associated to each distance. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of the probability density function. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data, considering SVM with linear kernel as classifier, a mislabeling process using the neighborhood of noisy samples and a noise level to control the number of errors in the data.

References

J. Du and Z. Cai. Modelling class noise with symmetric and asymmetric distributions. In Proc. 29th AAAI Conference on Artificial Intelligence, pages 2589-2595, 2015. url:https://dl.acm.org/doi/10.5555/2886521.2886681.

See Also

ugau_bor_ln, gaum_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- lap_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- lap_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Large-margin uniform label noise

Description

Introduction of Large-margin uniform label noise into a classification dataset.

Usage

## Default S3 method:
larm_uni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
larm_uni_ln(formula, data, ...)

Arguments

Details

Large-margin uniform label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, the samples are ordered according to their distance and (level·100)% of the most distant correctly classified samples to the decision boundary are selected to be mislabeled with a random different class.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data, considering SVM with linear kernel as classifier.

References

E. Amid, M. K. Warmuth, and S. Srinivasan. Two-temperature logistic regression based on the Tsallis divergence. In Proc. 22nd International Conference on Artificial Intelligence and Statistics, volume 89 of PMLR, pages 2388-2396, 2019. url:http://proceedings.mlr.press/v89/amid19a.html.

See Also

hubp_uni_ln, smu_cuni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- larm_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.3)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- larm_uni_ln(formula = Species ~ ., data = iris2D, level = 0.3)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Majority-class unidirectional label noise

Description

Introduction of Majority-class unidirectional label noise into a classification dataset.

Usage

## Default S3 method:
maj_udir_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
maj_udir_ln(formula, data, ...)

Arguments

Details

Let A be the majority class and B be the second majority class in the dataset. The Majority-class unidirectional label noise introduction model randomly selects (level·100)% of the samples of A and labels them as B.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data.

References

J. Li, Q. Zhu, Q. Wu, Z. Zhang, Y. Gong, Z. He, and F. Zhu. SMOTE- NaN-DE: Addressing the noisy and borderline examples problem in imbalanced classification by natural neighbors and differential evolution. Knowledge-Based Systems, 223:107056, 2021. doi:10.1016/j.knosys.2021.107056.

See Also

asy_def_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- maj_udir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- maj_udir_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Minority-driven bidirectional label noise

Description

Introduction of Minority-driven bidirectional label noise into a classification dataset.

Usage

## Default S3 method:
mind_bdir_ln(x, y, level, pos = 0.1, sortid = TRUE, ...)

## S3 method for class 'formula'
mind_bdir_ln(formula, data, ...)

Arguments

Details

Minority-driven bidirectional label noise randomly selects n = 2m·level samples in the dataset (with m the number of samples in the minority class), making sure that n·pos samples belong to the minority class and the rest to the majority class. Then, minority class samples are mislabeled as belonging to the majority class and majority class samples are mislabeled as belonging to the minority class. In case of ties determining minority and majority classes, a random class is chosen among them.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data.

References

A. Folleco, T. M. Khoshgoftaar, J. V. Hulse, and L. A. Bullard. Software quality modeling: The impact of class noise on the random forest classifier. In Proc. 2008 IEEE Congress on Evolutionary Computation, pages 3853–3859, 2008. doi:10.1109/CEC.2008.4631321.

See Also

fra_bdir_ln, irs_bdir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- mind_bdir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.5)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- mind_bdir_ln(formula = Species ~ ., data = iris2D, level = 0.5)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Minority-proportional uniform label noise

Description

Introduction of Minority-proportional uniform label noise into a classification dataset.

Usage

## Default S3 method:
minp_uni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
minp_uni_ln(formula, data, ...)

Arguments

Details

Given a dataset, assume the original class distribution of class i is pi and the distribution of the minority class is pm. Let level be the noise level, Minority-proportional uniform label noise introduces noise proportionally to different classes, where a sample with its label i has a probability (pm/pi)·level to be corrupted as another random class. That is, the least common class is used as the baseline for noise introduction.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

X. Zhu and X. Wu. Cost-guided class noise handling for effective cost-sensitive learning. In Proc. 4th IEEE International Conference on Data Mining, pages 297–304, 2004. doi:10.1109/ICDM.2004.10108.

See Also

asy_uni_ln, maj_udir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- minp_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- minp_uni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Misclassification prediction label noise

Description

Introduction of Misclassification prediction label noise into a classification dataset.

Usage

## Default S3 method:
mis_pre_ln(x, y, sortid = TRUE, ...)

## S3 method for class 'formula'
mis_pre_ln(formula, data, ...)

Arguments

Details

Misclassification prediction label noise creates a Multi-Layer Perceptron (MLP) model from the dataset and relabels each sample with the class predicted by the classifier.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

Q. Wang, B. Han, T. Liu, G. Niu, J. Yang, and C. Gong. Tackling instance-dependent label noise via a universal probabilistic model. In Proc. 35th AAAI Conference on Artificial Intelligence, pages 10183-10191, 2021. url:https://ojs.aaai.org/index.php/AAAI/article/view/17221.

See Also

smam_bor_ln, nlin_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- mis_pre_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)])

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- mis_pre_ln(formula = Species ~ ., data = iris2D)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Multiple-class unidirectional label noise

Description

Introduction of Multiple-class unidirectional label noise into a classification dataset.

Usage

## Default S3 method:
mulc_udir_ln(x, y, level, goal, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
mulc_udir_ln(formula, data, ...)

Arguments

Details

Multiple-class unidirectional label noise introduction model randomly selects (level·100)% of the samples of each class c with goal[c] != NA. Then, the labels c of these samples are replaced by the class indicated in goal[c]. The order of indices in goal is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

Q. Wang, B. Han, T. Liu, G. Niu, J. Yang, and C. Gong. Tackling instance-dependent label noise via a universal probabilistic model. In Proc. 35th AAAI Conference on Artificial Intelligence, pages 10183-10191, 2021. url:https://ojs.aaai.org/index.php/AAAI/article/view/17221.

See Also

minp_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- mulc_udir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1, 
                        goal = c(NA, 1, 2), order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- mulc_udir_ln(formula = Species ~ ., data = iris2D, level = 0.1, 
                        goal = c(NA, 1, 2), order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Neighborwise borderline label noise

Description

Introduction of Neighborwise borderline label noise into a classification dataset.

Usage

## Default S3 method:
nei_bor_ln(x, y, level, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
nei_bor_ln(formula, data, ...)

Arguments

Details

For each sample in the dataset, Neighborwise borderline label noise computes the ratio of two distances: the distance to its nearest neighbor from the same class and the distance to its nearest neighbor from another class. Then, these values are ordered in descending order and the first (level·100)% of them are used to determine the noisy samples. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering a mislabeling process using the neighborhood of noisy samples.

References

L. P. F. Garcia, J. Lehmann, A. C. P. L. F. de Carvalho, and A. C. Lorena. New label noise injection methods for the evaluation of noise filters. Knowledge-Based Systems, 163:693–704, 2019. doi:10.1016/j.knosys.2018.09.031.

See Also

ulap_bor_ln, lap_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- nei_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- nei_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Non-linearwise borderline label noise

Description

Introduction of Non-linearwise borderline label noise into a classification dataset.

Usage

## Default S3 method:
nlin_bor_ln(x, y, level, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
nlin_bor_ln(formula, data, ...)

Arguments

Details

Non-linearwise borderline label noise uses an SVM to induce the decision border in the dataset. Then, for each sample, its distance to the decision border is computed. Finally, the distances obtained are ordered in ascending order and the first (level·100)% of them are used to determine the noisy samples. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering a mislabeling process using the neighborhood of noisy samples.

References

L. P. F. Garcia, J. Lehmann, A. C. P. L. F. de Carvalho, and A. C. Lorena. New label noise injection methods for the evaluation of noise filters. Knowledge-Based Systems, 163:693–704, 2019. doi:10.1016/j.knosys.2018.09.031.

See Also

nei_bor_ln, ulap_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- nlin_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- nlin_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Type of noise introduced by a noise model

Description

Given the function name of a model, it returns the type of noise it introduces: label, attributes, or both.

Usage

noisetype(model)

Arguments

Value

A character with the type of noise model introduces. It can be cla for label noise, att for attribute noise or com for combined noise.

One-dimensional uniform label noise

Description

Introduction of One-dimensional uniform label noise into a classification dataset.

Usage

## Default S3 method:
oned_uni_ln(
  x,
  y,
  level,
  att,
  lower,
  upper,
  order = levels(y),
  sortid = TRUE,
  ...
)

## S3 method for class 'formula'
oned_uni_ln(formula, data, ...)

Arguments

Details

One-dimensional uniform label noise is based on the introduction of noise according to the values of the attribute att. Samples of class i with the attribute att falling between lower[i] and upper[i] have a probability level of being mislabeled. The labels of these samples are randomly replaced by other different ones within the set of class labels. The order of the class labels is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data, considering a noise level to control the number of errors in the data

References

N. Gornitz, A. Porbadnigk, A. Binder, C. Sannelli, M. L. Braun, K. Muller, and M. Kloft. Learning and evaluation in presence of non-i.i.d. label noise. In Proc. 17th International Conference on Artificial Intelligence and Statistics, volume 33 of PMLR, pages 293–302, 2014. url:https://proceedings.mlr.press/v33/gornitz14.html.

See Also

attm_uni_ln, qua_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- oned_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                    level = 0.5, att = 1, lower = c(1.5,2,6), upper = c(2,4,7))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- oned_uni_ln(formula = Species ~ ., data = iris2D, 
                    level = 0.5, att = 1, lower = c(1.5,2,6), upper = c(2,4,7))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Open-set ID/nearest-neighbor label noise

Description

Introduction of Open-set ID/nearest-neighbor label noise into a classification dataset.

Usage

## Default S3 method:
opes_idnn_ln(
  x,
  y,
  level,
  openset = c(1),
  order = levels(y),
  sortid = TRUE,
  ...
)

## S3 method for class 'formula'
opes_idnn_ln(formula, data, ...)

Arguments

Details

Open-set ID/nearest-neighbor label noise corrupts (level·100)% of the samples with classes in openset. Then, the labels of these samples are replaced by the label of the nearest sample of a different in-distribution class. The order of the class labels for the indices in openset is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

P. H. Seo, G. Kim, and B. Han. Combinatorial inference against label noise. In Advances in Neural Information Processing Systems, volume 32, pages 1171-1181, 2019. url:https://proceedings.neurips.cc/paper/2019/hash/0cb929eae7a499e50248a3a78f7acfc7-Abstract.html.

See Also

opes_idu_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- opes_idnn_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                      level = 0.4, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- opes_idnn_ln(formula = Species ~ ., data = iris2D, 
                      level = 0.4, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Open-set ID/uniform label noise

Description

Introduction of Open-set ID/uniform label noise into a classification dataset.

Usage

## Default S3 method:
opes_idu_ln(x, y, level, openset = c(1), order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
opes_idu_ln(formula, data, ...)

Arguments

Details

Open-set ID/uniform label noise corrupts (level·100)% of the samples with classes in openset. For each sample selected, a label from in-distribution classes is randomly chosen. The order of the class labels for the indices in openset is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

P. H. Seo, G. Kim, and B. Han. Combinatorial inference against label noise. In Advances in Neural Information Processing Systems, volume 32, pages 1171-1181, 2019. url:https://proceedings.neurips.cc/paper/2019/hash/0cb929eae7a499e50248a3a78f7acfc7-Abstract.html.

See Also

asy_spa_ln, mind_bdir_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- opes_idu_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                     level = 0.4, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- opes_idu_ln(formula = Species ~ ., data = iris2D, 
                     level = 0.4, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Pairwise bidirectional label noise

Description

Introduction of Pairwise bidirectional label noise into a classification dataset.

Usage

## Default S3 method:
pai_bdir_ln(x, y, level, pairs, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
pai_bdir_ln(formula, data, ...)

Arguments

Details

For each vector (c1, c2) in pairs, Pairwise bidirectional label noise randomly selects (level·100)% of the samples from class c1 in the dataset and (level·100)% of the samples from class c2. Then, c1 samples are mislabeled as belonging to c2 and c2 samples are mislabeled as belonging to c1. The order of the class labels is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

S. Fefilatyev, M. Shreve, K. Kramer, L. O. Hall, D. B. Goldgof, R. Kasturi, K. Daly, A. Remsen, and H. Bunke. Label-noise reduction with support vector machines. In Proc. 21st International Conference on Pattern Recognition, pages 3504-3508, 2012. url:https://ieeexplore.ieee.org/document/6460920/.

See Also

print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# create new class with some samples
class <- as.character(iris2D$Species)
class[iris2D$Petal.Length > 6] <- "newclass"
iris2D$Species <- as.factor(class)

# usage of the default method
set.seed(9)
outdef <- pai_bdir_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                      level = 0.1, pairs = list(c(1,2), c(3,4)), 
                      order = c("virginica", "setosa", "newclass", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- pai_bdir_ln(formula = Species ~ ., data = iris2D, 
                      level = 0.1, pairs = list(c(1,2), c(3,4)), 
                      order = c("virginica", "setosa", "newclass", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Plot function for class ndmodel

Description

Representation of the dataset contained in an object of class ndmodel after the application of a noise introduction model.

Usage

## S3 method for class 'ndmodel'
plot(x, ..., noise = NA, xvar = 1, yvar = 2, pca = FALSE)

Arguments

Details

This function performs a two-dimensional representation using the ggplot2 package of the dataset contained in the object x of class ndmodel. Each of the classes in the dataset (available in x$ynoise) is represented by a different color. There are two options to represent the input attributes of the samples on the x and y axes of the graph:

• If pca = FALSE, the values in the graph are taken from the current attribute values found in x$xnoise. In this case, xvar and yvar indicate the indices of the attributes to show in the x and y axes, respectively.

• If pca = TRUE, the values in the graph are taken after performing a PCA over x$xnoise. In this case, xvar and yvar indicate the index of the principal component according to the variance explained to show in the x and y axes, respectively.

Finally, the parameter noise is used to indicate which samples (noisy, clean or all) to show. Clean samples are represented by circles in the graph, while noisy samples are represented by crosses.

Value

An object of class ggplot and gg with the graph created using the ggplot2 package.

See Also

print.ndmodel, summary.ndmodel, sym_uni_ln, sym_cuni_ln, sym_uni_an

Examples

# load the dataset
data(iris)

# apply the noise introduction model
set.seed(9)
output <- sym_uni_ln(x = iris[,-ncol(iris)], y = iris[,ncol(iris)], level = 0.1)

# plots for all the samples, the clean samples and the noisy samples using PCA
plot(output, pca = TRUE)
plot(output, noise = FALSE, pca = TRUE)
plot(output, noise = TRUE, pca = TRUE)

# plots using the Petal.Length and Petal.Width variables
plot(output, xvar = 3, yvar = 4)
plot(output, noise = FALSE, xvar = 3, yvar = 4)
plot(output, noise = TRUE, xvar = 3, yvar = 4)

PMD-based confidence label noise

Description

Introduction of PMD-based confidence label noise into a classification dataset.

Usage

## Default S3 method:
pmd_con_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
pmd_con_ln(formula, data, ...)

Arguments

Details

PMD-based confidence label noise approximates the probability of noise using the confidence prediction of a neural network. These predictions are used to estimate the mislabeling probability and the most possible noisy class label for each sample. Finally, (level·100)% of the samples in the dataset are randomly selected to be mislabeled according to their values of probability computed.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

Y. Zhang, S. Zheng, P. Wu, M. Goswami, and C. Chen. Learning with feature-dependent label noise: A progressive approach. In Proc. 9th International Conference on Learning Representations, pages 1-13, 2021. url:https://openreview.net/forum?id=ZPa2SyGcbwh.

See Also

clu_vot_ln, sco_con_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- pmd_con_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- pmd_con_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Print function for class ndmodel

Description

This method displays the basic information about the noise introduction process contained in an object of class ndmodel.

Usage

## S3 method for class 'ndmodel'
print(x, ...)

Arguments

Details

This function presents the basic information of the noise introduction process and the resulting noisy dataset contained in the object x of class ndmodel. The information offered is as follows:

• the name of the noise introduction model.

• the parameters associated with the noise model.

• the number of noisy and clean samples in the dataset.

Value

This function does not return any value.

See Also

summary.ndmodel, plot.ndmodel, sym_uni_ln, sym_cuni_ln, sym_uni_an

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
print(outdef)

Print function for class sum.ndmodel

Description

Auxiliary function for printing information about the noise introduction process contained in an object of class sum.ndmodel.

Usage

## S3 method for class 'sum.ndmodel'
print(x, ...)

Arguments

Value

This function does not return any value.

Quadrant-based uniform label noise

Description

Introduction of Quadrant-based uniform label noise into a classification dataset.

Usage

## Default S3 method:
qua_uni_ln(x, y, level, att1 = 1, att2 = 2, sortid = TRUE, ...)

## S3 method for class 'formula'
qua_uni_ln(formula, data, ...)

Arguments

Details

For each sample, the probability of flipping its label is based on which quadrant (with respect to the attributes att1 and att2) the sample falls in. The probability of mislabeling for each quadrant is expressed with the argument level, whose length is equal to 4. Let m1 and m2 be the mean values of the domain of att1 and att2, respectively. Each quadrant is defined as follows: values <= m1 and <= m2 (first quadrant); values <= m1 and > m2 (second quadrant); values > m1 and <= m2 (third quadrant); and values > m1 and > m2 (fourth quadrant). Finally, the labels of these samples are randomly replaced by other different ones within the set of class labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

A. Ghosh, N. Manwani, and P. S. Sastry. Making risk minimization tolerant to label noise. Neurocomputing, 160:93-107, 2015. doi:10.1016/j.neucom.2014.09.081.

See Also

exps_cuni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- qua_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                       level = c(0.05, 0.15, 0.20, 0.4))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- qua_uni_ln(formula = Species ~ ., data = iris2D, 
                        level = c(0.05, 0.15, 0.20, 0.4))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Random numbers considering reference values

Description

Generate n random numbers following a uniform distribution between min and max. The values in ref can be chosen or not, according to original.

Usage

runif_replace(n, min, max, original, ref)

Arguments

Value

A double vector with the numbers generated.

Safe sample function

Description

Similar to standard sample function. Safe sample function considering the special case of an integer vector with only one element.

Usage

safe_sample(x, size, replace = FALSE, prob = NULL)

Arguments

Value

A vector with the elements chosen.

Sample considering reference values

Description

Similar to standard sample function. The values in ref can be chosen or not, according to original.

Usage

sample_replace(x, size, original, ref)

Arguments

Value

A vector with the elements chosen from x.

Score-based confidence label noise

Description

Introduction of Score-based confidence label noise into a classification dataset.

Usage

## Default S3 method:
sco_con_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sco_con_ln(formula, data, ...)

Arguments

Details

Score-based confidence label noise follows the intuition that hard samples are more likely to be mislabeled. Given the confidence per class of each sample, if it is predicted with a different class with a high probability, it means that it is hard to clearly distinguish the sample from this class. The confidence information is used to compute a mislabeling score for each sample and its potential noisy label. Finally, (level·100)% of the samples with the highest mislabeling scores are chosen as noisy.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

P. Chen, J. Ye, G. Chen, J. Zhao, and P. Heng. Beyond class-conditional assumption: A primary attempt to combat instance-dependent label noise. In Proc. 35th AAAI Conference on Artificial Intelligence, pages 11442-11450, 2021. url:https://ojs.aaai.org/index.php/AAAI/article/view/17363.

See Also

mis_pre_ln, smam_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sco_con_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sco_con_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Sigmoid-bounded uniform label noise

Description

Introduction of Sigmoid-bounded uniform label noise into a classification dataset.

Usage

## Default S3 method:
sigb_uni_ln(x, y, level, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sigb_uni_ln(formula, data, ...)

Arguments

Details

Sigmoid-bounded uniform label noise generates bounded instance-dependent and label-dependent label noise at random using a weight for each sample in the dataset to compute its noise probability through a sigmoid function. Note that this noise model considers the maximum noise level per class given by level, so the current noise level in each class may be lower than that specified. The order of the class labels is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data.

References

J. Cheng, T. Liu, K. Ramamohanarao, and D. Tao. Learning with bounded instance and label-dependent label noise. In Proc. 37th International Conference on Machine Learning, volume 119 of PMLR, pages 1789-1799, 2020. url:http://proceedings.mlr.press/v119/cheng20c.html.

See Also

larm_uni_ln, hubp_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sigb_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                      level = c(0.1, 0.2, 0.3))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sigb_uni_ln(formula = Species ~ ., data = iris2D, 
                      level = c(0.1, 0.2, 0.3))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Small-margin borderline label noise

Description

Introduction of Small-margin borderline label noise into a classification dataset.

Usage

## Default S3 method:
smam_bor_ln(x, y, level, k = 1, sortid = TRUE, ...)

## S3 method for class 'formula'
smam_bor_ln(formula, data, ...)

Arguments

Details

Small-margin borderline label noise uses an SVM to induce the decision border in the dataset. For each sample, its distance to the decision border is computed. Then, the samples are ordered according to their distance and (level·100)% of the closest correctly classified samples to the decision boundary are selected to be mislabeled. For each noisy sample, the majority class among its k-nearest neighbors of a different class is chosen as the new label.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References to multiclass data, considering SVM with linear kernel as classifier and a mislabeling process using the neighborhood of noisy samples.

References

E. Amid, M. K. Warmuth, and S. Srinivasan. Two-temperature logistic regression based on the Tsallis divergence. In Proc. 22nd International Conference on Artificial Intelligence and Statistics, volume 89 of PMLR, pages 2388-2396, 2019. url:http://proceedings.mlr.press/v89/amid19a.html.

See Also

nlin_bor_ln, nei_bor_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- smam_bor_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- smam_bor_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Smudge-based completely-uniform label noise

Description

Introduction of Smudge-based completely-uniform label noise into a classification dataset.

Usage

## Default S3 method:
smu_cuni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
smu_cuni_ln(formula, data, ...)

Arguments

Details

Smudge-based completely-uniform label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are randomly replaced by others within the set of class labels. An additional attribute smudge is included in the dataset with value equal to 1 in mislabeled samples and equal to 0 in clean samples.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

S. Thulasidasan, T. Bhattacharya, J. A. Bilmes, G. Chennupati, and J. Mohd-Yusof. Combating label noise in deep learning using abstention. In Proc. 36th International Conference on Machine Learning, volume 97 of PMLR, pages 6234-6243, 2019. url:http://proceedings.mlr.press/v97/thulasidasan19a.html.

See Also

oned_uni_ln, attm_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- smu_cuni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef, pca = TRUE)

# usage of the method for class formula
set.seed(9)
outfrm <- smu_cuni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Summary function for class ndmodel

Description

This method displays a summary containing information about the noise introduction process contained in an object of class ndmodel.

Usage

## S3 method for class 'ndmodel'
summary(object, ..., showid = FALSE)

Arguments

Details

This function presents a summary containing information of the noise introduction process and the resulting noisy dataset contained in the object object of class ndmodel. The information offered is as follows:

• the function call.

• the name of the noise introduction model.

• the parameters associated with the noise model.

• the number of noisy and clean samples in the dataset.

• the number of noisy samples per class/attribute.

• the number of clean samples per class/attribute.

• the indices of the noisy samples (if showid = TRUE).

Value

A list with the elements of object, including the showid argument.

See Also

print.ndmodel, plot.ndmodel, sym_uni_ln, sym_cuni_ln, sym_uni_an

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_uni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)

Symmetric adjacent label noise

Description

Introduction of Symmetric adjacent label noise into a classification dataset.

Usage

## Default S3 method:
sym_adj_ln(x, y, level, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sym_adj_ln(formula, data, ...)

Arguments

Details

Symmetric adjacent label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are replaced by a random adjacent class label according to order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

J. R. Cano, J. Luengo, and S. Garcia. Label noise filtering techniques to improve monotonic classification. Neurocomputing, 353:83-95, 2019. doi:10.1016/j.neucom.2018.05.131.

See Also

sym_dran_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_adj_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                    level = 0.1, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_adj_ln(formula = Species ~ ., data = iris2D, 
                     level = 0.1, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric center-based label noise

Description

Introduction of Symmetric center-based label noise into a classification dataset.

Usage

## Default S3 method:
sym_cen_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_cen_ln(formula, data, ...)

Arguments

Details

Symmetric center-based label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. The probability for chosing the noisy label is determined based on the distance between class centers. Thus, the mislabeling probability between classes increases as the distance between their centers decreases. This model is consistent with the intuition that samples in similar classes are more likely to be mislabeled. Besides, the model also allows mislabeling data in dissimilar classes with a relatively small probability, which corresponds to label noise caused by random errors.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

X. Pu and C. Li. Probabilistic information-theoretic discriminant analysis for industrial label-noise fault diagnosis. IEEE Transactions on Industrial Informatics, 17(4):2664-2674, 2021. doi:10.1109/TII.2020.3001335.

See Also

glev_uni_ln, sym_hienc_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_cen_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_cen_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric confusion label noise

Description

Introduction of Symmetric confusion label noise into a classification dataset.

Usage

## Default S3 method:
sym_con_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_con_ln(formula, data, ...)

Arguments

Details

Symmetric confusion label noise considers that the mislabeling probability for each class is level. It obtains the confusion matrix from the dataset, which is row-normalized to estimate the transition matrix and get the probability of selecting each class when noise occurs.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, considering C5.0 as classifier.

References

D. Ortego, E. Arazo, P. Albert, N. E. O’Connor, and K. McGuinness. Towards robust learning with different label noise distributions. In Proc. 25th International Conference on Pattern Recognition, pages 7020-7027, 2020. doi:10.1109/ICPR48806.2021.9412747.

See Also

sym_cen_ln, glev_uni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_con_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_con_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric completely-uniform attribute noise

Description

Introduction of Symmetric completely-uniform attribute noise into a classification dataset.

Usage

## Default S3 method:
sym_cuni_an(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_cuni_an(formula, data, ...)

Arguments

Details

Symmetric completely-uniform attribute noise corrupts (level·100)% of the values of each attribute in the dataset. In order to corrupt an attribute A, (level·100)% of the samples in the dataset are randomly chosen. Then, their values for A are replaced by random ones from the domain of the attribute. Note that the original attribute value of a sample can be chosen as noisy and the actual percentage of noise in the dataset can be lower than the theoretical noise level.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References, only considering attribute noise introduction.

References

C. Teng. Polishing blemishes: Issues in data correction. IEEE Intelligent Systems, 19(2):34-39, 2004. doi:10.1109/MIS.2004.1274909.

See Also

sym_uni_an, sym_cuni_cn, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_cuni_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_cuni_an(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric completely-uniform combined noise

Description

Introduction of Symmetric completely-uniform combined noise into a classification dataset.

Usage

## Default S3 method:
sym_cuni_cn(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_cuni_cn(formula, data, ...)

Arguments

Details

Symmetric completely-uniform combined noise corrupts (level·100)% of the values of each attribute in the dataset. In order to corrupt an attribute A, (level·100)% of the samples in the dataset are randomly chosen. Then, their values for A are replaced by random ones from the domain of the attribute.

Additionally, this noise model also selects (level·100)% of the samples in the dataset with independence of their class. The labels of these samples are randomly replaced by other ones within the set of class labels.

Note that, for both attributes and class labels, the original value of a sample can be chosen as noisy and the actual percentage of noise in the dataset can be lower than the theoretical noise level.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

C. Teng. Polishing blemishes: Issues in data correction. IEEE Intelligent Systems, 19(2):34-39, 2004. doi:10.1109/MIS.2004.1274909.

See Also

uncs_guni_cn, sym_cuni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_cuni_cn(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_cuni_cn(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric completely-uniform label noise

Description

Introduction of Symmetric completely-uniform label noise into a classification dataset.

Usage

## Default S3 method:
sym_cuni_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_cuni_ln(formula, data, ...)

Arguments

Details

Symmetric completely-uniform label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are randomly replaced by others within the set of class labels. Note that this model can choose the original label of a sample as noisy.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

A. Ghosh and A. S. Lan. Contrastive learning improves model robustness under label noise. In Proc. 2021 IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2703-2708, 2021. doi:10.1109/CVPRW53098.2021.00304.

See Also

sym_uni_ln, sym_cuni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_cuni_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_cuni_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric double-default label noise

Description

Introduction of Symmetric double-default label noise into a classification dataset.

Usage

## Default S3 method:
sym_ddef_ln(
  x,
  y,
  level,
  def1 = 1,
  def2 = 2,
  order = levels(y),
  sortid = TRUE,
  ...
)

## S3 method for class 'formula'
sym_ddef_ln(formula, data, ...)

Arguments

Details

Symmetric double-default label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are replaced by one of two fixed labels (def1 or def2) within the set of class labels. The indices def1 and def2 are taken according to the order given by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

B. Han, J. Yao, G. Niu, M. Zhou, I. W. Tsang, Y. Zhang, and M. Sugiyama. Masking: A new perspective of noisy supervision. In Advances in Neural Information Processing Systems, volume 31, pages 5841-5851, 2018. url:https://proceedings.neurips.cc/paper/2018/hash/aee92f16efd522b9326c25cc3237ac15-Abstract.html.

See Also

sym_exc_ln, sym_cuni_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_ddef_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)],
                      level = 0.1, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_ddef_ln(formula = Species ~ ., data = iris2D,
                      level = 0.1, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric default label noise

Description

Introduction of Symmetric default label noise into a classification dataset.

Usage

## Default S3 method:
sym_def_ln(x, y, level, def = 1, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sym_def_ln(formula, data, ...)

Arguments

Details

Symmetric default label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are replaced by a fixed label (def) within the set of class labels. The index def is taken according to the order given by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

M. Ren, W. Zeng, B. Yang, and R. Urtasun. Learning to reweight examples for robust deep learning. In Proc. 35th International Conference on Machine Learning, volume 80 of PMLR, pages 4331-4340, 2018. url:http://proceedings.mlr.press/v80/ren18a.html.

See Also

sym_ddef_ln, sym_exc_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_def_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                     level = 0.1, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_def_ln(formula = Species ~ ., data = iris2D,
                     level = 0.1, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric diametrical label noise

Description

Introduction of Symmetric diametrical label noise into a classification dataset.

Usage

## Default S3 method:
sym_dia_ln(x, y, level, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sym_dia_ln(formula, data, ...)

Arguments

Details

Symmetric diametrical label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. In this model, diametrical (opposite) classes are more likely to have their labels mixed. The probability of mislabel a sample of class i as belonging to class j is computed as dij/S, where dij = abs(i-j) and S is the sum of distances to class i. The order of the classes is determined by order.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

R. C. Prati, J. Luengo, and F. Herrera. Emerging topics and challenges of learning from noisy data in nonstandard classification: a survey beyond binary class noise. Knowledge and Information Systems, 60(1):63–97, 2019. doi:10.1007/s10115-018-1244-4.

See Also

sym_pes_ln, sym_opt_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_dia_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], 
                     level = 0.1, order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_dia_ln(formula = Species ~ ., data = iris2D, 
                     level = 0.1, order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric double-random label noise

Description

Introduction of Symmetric double-random label noise into a classification dataset.

Usage

## Default S3 method:
sym_dran_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_dran_ln(formula, data, ...)

Arguments

Details

Symmetric double-random label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, each of the original class labels is flipped to one between two other random labels.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

A. Ghosh and A. S. Lan. Do we really need gold samples for sample weighting under label noise? In Proc. 2021 IEEE Winter Conference on Applications of Computer Vision, pages 3921-3930, 2021. doi:10.1109/WACV48630.2021.00397.

See Also

sym_hie_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_dran_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_dran_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric end-directed attribute noise

Description

Introduction of Symmetric end-directed attribute noise into a classification dataset.

Usage

## Default S3 method:
sym_end_an(x, y, level, scale = 0.2, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_end_an(formula, data, ...)

Arguments

Details

For each attribute A, Symmetric end-directed attribute noise computes a value k = scale·max(A). Then, it chooses (level·100)% of the values of that attribute. For each value, it applies the following procedure:

• If the value is less than the median of the attribute, the value transforms into adding k to the maximum of the attribute A.

• If the value is greater than the median of the attribute, the value transforms into subtracting k from the minimum of the attribute A.

• If the value matches the median, one of the two previous alternatives is chosen.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

T. M. Khoshgoftaar and J. V. Hulse. Empirical case studies in attribute noise detection. IEEE Transactions on Systems, Man and Cybernetics Part C: Applications and Reviews, 39(4):379-388, 2009. doi:10.1109/TSMCC.2009.2013815.

See Also

sym_sgau_an, symd_gau_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_end_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_end_an(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric exchange label noise

Description

Introduction of Symmetric exchange label noise into a classification dataset.

Usage

## Default S3 method:
sym_exc_ln(x, y, level, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_exc_ln(formula, data, ...)

Arguments

Details

Symmetric exchange label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. These samples are divided into two groups: A and B. Then, each sample of group A is labeled with the label of a sample of group B and vice versa.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

J. Schneider, J. P. Handali, and J. vom Brocke. Increasing trust in (big) data analytics. In Proc. 2018 Advanced Information Systems Engineering Workshops, volume 316 of LNBIP, pages 70-84, 2018. doi:10.1007/978-3-319-92898-2_6.

See Also

sym_cuni_ln, sym_cuni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_exc_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_exc_ln(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric Gaussian attribute noise

Description

Introduction of Symmetric Gaussian attribute noise into a classification dataset.

Usage

## Default S3 method:
sym_gau_an(x, y, level, k = 0.2, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_gau_an(formula, data, ...)

Arguments

Details

Symmetric Gaussian attribute noise corrupts (level·100)% of the values of each attribute in the dataset. In order to corrupt an attribute A, (level·100)% of the samples in the dataset are chosen. Then, their values for A are corrupted adding a random value that follows a Gaussian distribution of mean = 0 and standard deviation = (max-min)·k, being max and min the limits of the attribute domain. For nominal attributes, a random value is chosen.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

J. A. Sáez, M. Galar, J. Luengo, and F. Herrera. Analyzing the presence of noise in multi-class problems: alleviating its influence with the one-vs-one decomposition. Knowledge and Information Systems, 38(1):179-206, 2014. doi:10.1007/s10115-012-0570-1.

See Also

sym_int_an, symd_uni_an, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_gau_an(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1)

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_gau_an(formula = Species ~ ., data = iris2D, level = 0.1)

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric hierarchical label noise

Description

Introduction of Symmetric hierarchical label noise into a classification dataset.

Usage

## Default S3 method:
sym_hie_ln(x, y, level, group, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sym_hie_ln(formula, data, ...)

Arguments

Details

Symmetric hierarchical label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are randomly replaced by other ones within the set of class labels related to them (given by the argument group). The indices in group are taken according to the order given by order. Note that if a class does not belong to any superclass, it may be mislabeled as any other class.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

D. Hendrycks, M. Mazeika, D. Wilson, and K. Gimpel. Using trusted data to train deep networks on labels corrupted by severe noise. In Advances in Neural Information Processing Systems, volume 31, pages 10477-10486, 2018. url:https://proceedings.neurips.cc/paper/2018/hash/ad554d8c3b06d6b97ee76a2448bd7913-Abstract.html.

See Also

sym_uni_ln, sym_def_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method: a superclass with labels of indices 1 and 2
set.seed(9)
outdef <- sym_hie_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1, 
                       group = list(c(1,2)), order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_hie_ln(formula = Species ~ ., data = iris2D, level = 0.1,
                       group = list(c(1,2)), order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric hierarchical/next-class label noise

Description

Introduction of Symmetric hierarchical/next-class label noise into a classification dataset.

Usage

## Default S3 method:
sym_hienc_ln(x, y, level, group, order = levels(y), sortid = TRUE, ...)

## S3 method for class 'formula'
sym_hienc_ln(formula, data, ...)

Arguments

Details

Symmetric hierarchical/next-class label noise randomly selects (level·100)% of the samples in the dataset with independence of their class. Then, the labels of these samples are replaced by the next class within the set of class labels related to them (given by the argument group). The indices in group are taken according to the order given by order. Note that if a class does not belong to any superclass, it may be mislabeled as any other class.

Value

An object of class ndmodel with elements:

Note

Noise model adapted from the papers in References.

References

T. Kaneko, Y. Ushiku, and T. Harada. Label-noise robust generative adversarial networks. In Proc. 2019 IEEE Conference on Computer Vision and Pattern Recognition, pages 2462-2471, 2019. doi:10.1109/CVPR.2019.00257.

See Also

sym_nexc_ln, sym_dia_ln, print.ndmodel, summary.ndmodel, plot.ndmodel

Examples

# load the dataset
data(iris2D)

# usage of the default method
set.seed(9)
outdef <- sym_hienc_ln(x = iris2D[,-ncol(iris2D)], y = iris2D[,ncol(iris2D)], level = 0.1, 
                       group = list(c(1,2)), order = c("virginica", "setosa", "versicolor"))

# show results
summary(outdef, showid = TRUE)
plot(outdef)

# usage of the method for class formula
set.seed(9)
outfrm <- sym_hienc_ln(formula = Species ~ ., data = iris2D, level = 0.1,
                       group = list(c(1,2)), order = c("virginica", "setosa", "versicolor"))

# check the match of noisy indices
identical(outdef$idnoise, outfrm$idnoise)

Symmetric interval-based attribute noise

Description

Introduction of Symmetric interval-based attribute noise into a classification dataset.

Usage

## Default S3 method:
sym_int_an(x, y, level, nbins = 10, sortid = TRUE, ...)

## S3 method for class 'formula'
sym_int_an(formula, data, ...)

Arguments

Details

Symmetric interval-based attribute noise corrupts (level·100)% of the values of each attribute in the dataset. In order to corrupt an attribute A, (level·100)% of the samples in the dataset are selected. To corrupt numeric attributes, the attribute is split into nbins equal-frequency intervals, one of its closest intervals is chosen and a random value within the interval is picked out as noisy. For nominal attributes, a random value within the domain is chosen.

Value