Heterogeneous Multi-task Feature Learning

Description

HMTL package implements the block-wise sparse estimation by grouping the coefficients of related predictors across multiple tasks. The tasks can be either regression, Huber regression, adaptive Huber regression, and logistic regression, which provide a wide variety of data types for the integration. The robust methods, such as the Huber regression and adaptive Huber regression, can deal with outlier contamination based on Sun, Q., Zhou, W.-X. and Fan, J. (2020), and Wang, L., Zheng, C., Zhou, W. and Zhou, W.-X. (2021). The model selection applies the modified form of Bayesian information criterion to measure the mdoel performance, which has similar formulation as previous work developed by Gao, X., and Carroll, R. J., (2017).

Details

In the context of multi-task learning, there are K different data sets obtained from K related sources. The data sets can be modeled by different types of learning tasks based on the data distributions. Let the candidate features be denoted as \{M_1,M_2,...,M_j,...,M_p \}. When the integrated data sets have different measurements, we assume the predictors to share some similarities. For example, the jth predictors collected as M_j = (X_{1j}, X_{2j}, \cdots, X_{Kj}) in the table below represent the same type of feature in all related studies. In some cases, the tasks can share same set of predictor, then X_{1j} = X_{2j} = \cdots = X_{Kj}.

The coefficients can be grouped as the vector \theta^{(j)} for the feature M_j.

The heterogeneous multi-task feature learning HMTL can select significant features through the overall objective function:

Q(\theta)= \mathcal{L}(\theta) + \mathcal{R}(\theta).

The loss function is defined as \mathcal{L}(\theta) = \sum_{k=1}^K w_k \ell_k(\theta_k), which can be the composite quasi-likelihood or the composite form of (adaptive) Huber loss with additional robustification parameter \tau_k. The penalty function is the mixed \ell_{2,1} regularization, such that \mathcal{R}(\theta) = \lambda \sum_{j=1}^p (\sum_{k=1}^K \theta_{kj}^2)^{1/2}.

This package also contains functions to provide the Bayesian information criterion:

BIC(s) = 2\mathcal{L}_s(\hat{\theta}) + d_s^{*} \gamma_n

with \mathcal{L}_s(\hat{\theta}) denoting the composite quasi-likelihood or adaptive Huber loss, d_s^{*} measuring the model complexity and \gamma_n being the penalty on the model complexity.

In this package, the function MTL_reg deals with regression tasks, which can be outlier contaminated. The function MTL_class is applied to model multiple classification tasks, and the function MTL_hetero can integrate different types of tasks together.

Author(s)

Yuan Zhong, Wei Xu, and Xin Gao

Maintainer: Yuan Zhong <aqua.zhong@gmail.com>

References

Zhong, Y., Xu, W., and Gao X., (2023) Heterogeneous multi-task feature learning with mixed \ell_{2,1} regularization. Submitted

Zhong, Y., Xu, W., and Gao X., (2023) Robust Multi-task Feature Learning. Submitted

Gao, X., and Carroll, R. J., (2017) Data integration with high dimensionality. Biometrika, 104, 2, pp. 251-272

Huber, P. J. (1964). Robust estimation of a location parameter. Ann. Math. Statist., 35, 73–101.

Sun, Q., Zhou, W.-X. and Fan, J. (2020). Adaptive Huber regression. J. Amer. Statist. Assoc., 115, 254-265.

Wang, L., Zheng, C., Zhou, W. and Zhou, W.-X. (2021). A new principle for tuning-free Huber regression. Stat. Sinica, 31, 2153-2177.

Multiple Classification Task Feature Learning

Description

MTL_class conducts multi-tasks feature learning to the learning tasks with binary response variables, namely logistic regression. The penalty function applies a mixed \ell_{2,1} norm to combine regression coefficients of predictor shared across all tasks.

Usage

MTL_class(
  y,
  x,
  lambda,
  Kn,
  p,
  n,
  beta = 0.1,
  import_w = 1,
  tol = 0.05,
  max_iter = 100,
  Complete = "True",
  diagnostics = FALSE,
  gamma = 1,
  alpha = 1
)

Arguments

Value

A list including the following terms will be returned:

beta

A p by K matrix of estimated sparse parameters.

Task type

The models used in each task.

Task weights

The weights assigned to each task.

Selected_List

The index of non-zero parameters.

If 'diagnostics = TRUE', the following terms will be returned:

Bayesian_Information

Table of the information criterion: Composite likelihood, Degree of freedom, and (peudo or robust) Bayesian informtion criterion.

Class_Perform

Table of the model performance for classification tasks: the area under ROC curve (AUC), and the deviance (DEV) estimated by 'glm'.

Residuals

The residuals for all tasks.

References

Zhong, Y., Xu, W., and Gao X., (2023) Heterogeneous multi-task feature learning with mixed \ell_{2,1} regularization. Submitted

Examples

x_class <- list(mockdata1[[3]],mockdata1[[4]])
y_class <- list(mockdata2[[3]],mockdata2[[4]])
model <- MTL_class(y_class,x_class, lambda = 2/11 , Kn = 2, p=500,
                  n = 250 ,gamma = 1, Complete = FALSE, diagnostics = TRUE, alpha = 2)
# Selected non-zero coefficients
model$beta[model$Selected_List,]
# Estimated Pseudo-BIC
model$Bayesian_Information
# Classification accuracy
model$Class_Perform

Heterogeneous Multi-task Feature Learning

Description

MTL_hetero conducts multi-tasks feature learning to different types of learning tasks, including linear regression, Huber regression, adaptive Huber, and logistic regression. The penalty function applies a mixed \ell_{2,1} norm to combine regression coefficients of predictor shared across all tasks.

Usage

MTL_hetero(
  y,
  x,
  lambda,
  Kn,
  p,
  n,
  beta = 0.1,
  tau = 1.45,
  Cont_Model = "adaptive Huber",
  import_w = 1,
  tol = 0.05,
  max_iter = 100,
  Complete = "True",
  diagnostics = FALSE,
  gamma = 1,
  alpha = 1
)

Arguments

Value

A list including the following terms will be returned:

beta

A p by K matrix of estimated sparse parameters.

Task type

The models used in each task.

Task weights

The weights assigned to each task.

Selected_List

The index of non-zero parameters.

If 'diagnostics = TRUE', the following terms will be returned:

Bayesian_Information

Table of the information criterion: Composite likelihood, Degree of freedom, and (peudo or robust) Bayesian informtion criterion.

Reg_Error

Table of the model performance for (Huber) regressions: the mean square error (MSE), and the mean absolute error (MAE).

Class_Perform

Table of the model performance for classification tasks: the area under ROC curve (AUC), and the deviance (DEV) estimated by 'glm'.

Residuals

The residuals for all tasks.

Note

When penalty parameter is too small, the estimated coefficients may have p \ge n. The algorithm can provide the estimated values, but the diagnostics's results will not be given.

References

Zhong, Y., Xu, W., and Gao X., (2023) Heterogeneous multi-task feature learning with mixed \ell_{2,1} regularization. Submitted

Examples

model <- MTL_hetero(mockdata2,mockdata1, lambda =2.5, Kn = c(2,2), p=500,n = c(500,250,250,250),
                    gamma = 2, Complete = FALSE, diagnostics = TRUE, alpha = 2)
# Selected non-zero coefficients
model$beta[model$Selected_List,]
# Estimated Pseudo-BIC
model$Bayesian_Information
# Regression error
model$Reg_Error
# Classification accuracy
model$Class_Perform

Robust Multi-Task Feature Learning

Description

MTL_reg conducts multi-tasks feature learning to the learning tasks with continous response variables, such as the linear regression, Huber regression, adaptive Huber. The adaptive Huber method is based on Sun, Q., Zhou, W.-X. and Fan, J. (2020) and Wang, L., Zheng, C., Zhou, W. and Zhou, W.-X. (2021). The penalty function applies a mixed \ell_{2,1} norm to combine regression coefficients of predictor shared across all tasks. The Huber regression and adaptive Huber regression need the robustification parameter \tau_k to strike a balance between the unbiasedness and robustness, and the adaptive method can determine this parameter by a tuning-free principle.

Usage

MTL_reg(
  y,
  x,
  lambda,
  Kn,
  p,
  n,
  beta = 0.1,
  tau = 1.45,
  Cont_Model = "adaptive Huber",
  import_w = 1,
  tol = 0.05,
  max_iter = 100,
  Complete = "True",
  diagnostics = FALSE,
  gamma = 1,
  alpha = 1
)

Arguments

Value

A list including the following terms will be returned:

beta

A p by K matrix of estimated sparse parameters.

Task type

The models used in each task.

Task weights

The weights assigned to each task.

Selected_List

The index of non-zero parameters.

If 'diagnostics = TRUE', the following terms will be returned:

Bayesian_Information

Table of the information criterion: Composite likelihood, Degree of freedom, and (peudo or robust) Bayesian informtion criterion.

Reg_Error

Table of the model performance for (Huber) regressions: the mean square error (MSE), and the mean absolute error (MAE).

Residuals

The residuals for all tasks.

References

Huber, P. J. (1964). Robust estimation of a location parameter. Ann. Math. Statist., 35, 73–101.

Sun, Q., Zhou, W.-X. and Fan, J. (2020). Adaptive Huber regression. J. Amer. Statist. Assoc., 115, 254-265.

Wang, L., Zheng, C., Zhou, W. and Zhou, W.-X. (2021). A new principle for tuning-free Huber regression. Stat. Sinica, 31, 2153-2177.

Zhong, Y., Xu, W., and Gao X., (2023) Robust Multi-task Feature Learning. Submitted

Examples

x_reg <- list(mockdata1[[1]],mockdata1[[2]])
y_reg <- list(mockdata2[[1]],mockdata2[[2]])
model <- MTL_reg(y_reg,x_reg, lambda = 2.5  , Kn = 2, p=500,
                n = c(500,250 ),gamma = 2, Complete = FALSE, diagnostics = TRUE, alpha = 2)

# Selected non-zero coefficients
model$beta[model$Selected_List,]
# Estimated Pseudo-BIC
model$Bayesian_Information
# Regression error
model$Reg_Error

Model Selection for Multi-task Feature Learning based on Bayesian Information Criterion (BIC)

Description

Selection_HMTL can be used to search the optimal candidate model based on Bayesian Information Criterion (BIC).

Usage

Selection_HMTL(
  y,
  x,
  lambda,
  Kn,
  p,
  n,
  beta = 0.1,
  tau = 1.45,
  Cont_Model = "adaptive Huber",
  type = "Heterogeneity",
  import_w = 1,
  tol = 0.05,
  max_iter = 100,
  Complete = "True",
  diagnostics = FALSE,
  gamma = 1,
  alpha = 1
)

Arguments

Details

The Bayesian information criterion is given by

BIC(s) = 2\mathcal{L}_s(\hat{\theta}) + d_s^{*} \gamma_n,

where \hat{\theta} is the estimated coefficients and s is denoted the selected support set of \hat{\theta}. In addition, \mathcal{L}_s(\hat{\theta}) denoted the estimated composite quasi-likelihood or adaptive Huber loss evaluated as the \hat{\theta}. The degree of freedom d_s^{*} can measure the model complexity, which is estimated by tr(H_s^{-1}(\hat{\theta}) J_s(\hat{\theta}) ). The sensitivity matrix and specificity matrix can be given by H(\theta) = E(\nabla^2 \mathcal{L}( {\theta})) and J(\theta) = -Cov(\nabla \mathcal{L}( {\theta})). The penalty term \gamma_n can be defined by users.

Value

A table of Bayesian Information Criterion (BIC)

lambda

A list of penalty parameters.

Compo_likelihood

Sum of empirical loss functions estimated based on the selected parameters .

Degree of freedom

Penalty component based on the selected parameters.

Info criterion

Bayesian Information Criterion (BIC): robust BIC or pseudo BIC.

References

Y. Zhong, W. Xu, and X. Gao (2023) Robust Multi-task Feature Learning. Submitted

Gao, X and Carroll, R. J. (2017) Data integration with high dimensionality. Biometrika, 104, 2, pp. 251-272

Examples

lambda <- c(1.2, 2.0, 2.3, 2.8)
cv.mtl <- Selection_HMTL(mockdata2,mockdata1, lambda =lambda, Kn = c(2,2), p=500,
               n = c(500,250,250,250),gamma = 2, Complete = FALSE, diagnostics = TRUE, alpha = 2)
plot_HMTL(cv.mtl)
cv.mtl$Selection_Results

Mock Gene Data

Description

This data set is a mock version of two related research outcomes.

Usage

data("mockdata1")
data("mockdata2")

Format

The mockdata1 contains all predictor variables for four data sets, and the mockdata2 is the list of four columns of response variables. The sample sizes are n_1 = 500, n_2 = 250, n_3 = 250, and n_4 = 250.

The response varibles are heterogeneous, such that the first two columns are continuous data and last two are binary data. The predictors of each data matrix in mockdata1 have 500 columns of variables, and the columns in different data matrices are matched. For example, the first column in the first data matrix represents the same type of feature as the first columns in other three data matrices. Among all canadidate predictors, the response variables are set to be associated with the first 22 columns, and the remaining columns are not important predictors.

Details

This data set is used as an example to implement functions in the package.

Plot diagram of the information criterion vs. penalty parameters

Description

Plot a diagram to illustrate the change of Bayesian information criterion with different penalty paameters for model selection

Usage

plot_HMTL(x)

Arguments

Value

X axis represents the value of penalty parameters, and y axis represents the estimated values of the composite likelihood, degree of freedom for model complexity, and the (robust) Bayesian information criterion.