Group Sparse Multi-task Learning Method And Its Application

When building a machine learning model,there exit multiple correlated tasks to be solved,such as classifying the images under various lighting,background,and shooting angles condition.Hence multi-task learning is proposed to address the above problems.We aim to create more robust and generalized model by utilizing task correlations,potential and useful additional information.This paper mainly studies the group sparse lasso multi-task learning method,given the input and output data of known tasks,this method improves the prediction performance of the model by using samples of other related tasks,which overcomes the difficulties of limited sample size of a specific task.At the same time,the limiting model based group sparse regularization is introduced to extracting the features.This regularization can make full use of the effectiveness of feature screening to filter the desired features.The procedure can be described as follows: first,The 881 dimensional features of samples can be extracted from the dataset by using the PubChem molecular fingerprint method.then,the task correlations are incorporated into group sparse regularization to construct the group sparsity multi-task learning model.Finally,the feature selection method is employed to filter out desired features for a specific application.The final experimental results show that,In comparison with the single-task learning,the deep multi-task learning method,and other regularization multi-task learning method,our proposed method can achieve better performance in terms of prediction accuracy and feature selection.