Multi-task ordinal regression with task weight discovery

Abstract

Ordinal regression (OR) deals with the classification problems that the classes are ranked in order. At present, most OR approaches are designed for individual tasks, the research on multi-task OR is limited. These multi-task OR approaches assume that different tasks have the same relatedness and contribute equally to the overall model. However, in practice, different tasks may have distinct relatedness to the overall model. If they are treated equally, the performance of the overall model may be restricted. In this paper, we propose a novel multi-task OR approach with task weight discovery (MORTD). We assign each task a weight that indicates its relatedness to the overall model. Based on the task weights, a maximum margin multi-task OR model is constructed. Then, we adopt a heuristic framework to construct the multi-task OR classifier and update the task weights alternately. In this framework, the dual coordinate descent method is adapted to train the multi-task OR classifier efficiently. In real-world OR applications, the relatedness of multiple tasks may not be exactly the same. The contribution of MORTD is that it can discover the weights of tasks to yield a more precise classification model. Substantial experiments on real-life OR datasets illustrate that compared to the existing multi-task OR methods, MORTD is able to deliver higher classification accuracy and meanwhile needs less training time.