Adaptive dual graph regularization for clustered multi-task learning

The key challenge of multi -task learning is how to exploit the structure across all tasks. In practice, relevant tasks are partially associated with similar meaningful feature subgroups, which implies an overlapped taskfeature co-cluster structure. Besides discovering relationships at the task level, collaboratively identifying relevant meaningful structure relationship among features is beneficial to properly capture the structure of tasks. Toward this aim, we propose a clustered multi -task learning approach that collaboratively learns the cluster structure for both task and feature level effects. Specifically, an adaptive dual graph regularization, which respectively formulates the similarity of tasks and features, is introduced to guide the learning process to discover task -feature co-cluster relationship in a flexible way. Additionally, without any prior knowledge, the similarity weight of dual graph regularization can be automatically inferred through adaptive graph learning during model training. Experimental studies validate the effectiveness of our approach in terms of improving predictive performance and capturing clear cluster structure among tasks. The source code of the proposed method is available at GitHub: https://github.com/CLiu272/AdualGraph.