DMTMV: A Unified Learning Framework for Deep Multi-task Multi-view Learning

As the development of data collection techniques, complicated objects are described with more than one aspects as well as possess multiple concepts, so as many data mining approaches face the issues of dual-heterogeneity, i.e., feature heterogeneity and task heterogeneity. Traditional multi-task learning methods and multi-view learning methods may be not optimal for such a complicated learning problem since they only capture one type of heterogeneity. Then some works concentrate on a new direction where there are multiple related tasks with multi-view data. However, most existing MTMV methods focus on proposing linear models for fitting the specific application requirements while they are not suitable for common large-scale real-world problems in real environments. In this paper, we propose a unified learning framework for a deep multi-task multi-view neural network. In our approach, there are three kinds of networks called shared feature network, specific feature network and task network, each of which focus on the feature heterogeneity, unified feature representations, and task heterogeneity, respectively. Meanwhile, we employ a layer-by-layer regularization strategy for learning the relationships between tasks in multi-task multi-view learning. Moreover, the DMTMV method is naturally convenient for multi-class heterogeneous tasks as well. Finally, experiments on four real-world datasets successfully show that the proposed framework can significantly improve the prediction performance in multi-task multi-view learning while it can also discover inherent relationships among different tasks.