Additive Model Based Multi-task Learning And Applications

Multi-task learning is a significant branch in machine learning and data mining communities, which seeks to learn multiple tasks simultaneously and hence improves the generalization performance, for the scenarios that samples are draw from different distributions and are extremely insufficient. Furthermore, with the increase of complex-ity of the relationship structure among tasks in real applications, the additive multi-task learning approaches become an important tool.This paper in concerned with additive multi-task learning, and the works we con-ducted are as follows:1. We discussed an important aspect in multi-tasks learning-to extract the rela-tionships between tasks and features, and proposed a novel co-clustering based multi-task learning CoCTML, which utilizes the relationship between tasks and features, resulting co-clusters which contains task nodes and feature nodes simul-taneously. The CoCTML approach enforces the block-sparsity for the weight matrix, and we proposed a proximal algorithm based on PALM to solve the prob-lem.2. Then we proposed another novel additive multi-task learning method-RDMTR-for the trajectory regression problem for traffic data, which is able to capture the dynamic temporal changes while enforcing the spatial and temporal smoothness. The RDMTR could extract the extra cost caused by the traffic jam in rush hours. Later, we designed a proximal algorithm to solve the convex problem.We conducted the experiments both on synthetic and real data sets. It should be mentioned that both the CoCTML and RDMTR outperforms the other approaches in comparison, which verifies the effectiveness of CoCTML and RDMTR. Moreover, we analysed the weight matrix which represents the extra cost in RDMTR, finding that RDMTR is capable to recover the peak periods for traffic scenarios, which verifies the success of additive multi-task learning techniques in the problem of trajectory regres-sion.