A Regression Model Tree Algorithm by Multi-task Learning

To tackle the small sample size problem and difficulty reflecting a global effect in a model tree, this study proposes a regression model tree algorithm that employs multi-task learning. By applying multi-task learning considering the relatedness among terminal nodes when estimating a regression model tree, the proposed method prevents overfitting in a node with limited training data and considers both local and global effects. In addition, the multi-task learning algorithm used in the proposed method, interprets the relationship between different nodes. Experimental results on synthetic and real datasets demonstrate that the proposed method improves the prediction performance over baseline methods, particularly in small-sized nodes.