Towards Interpretable Multi-task Learning Using Bilevel Programming

Interpretable Multi-Task Learning can be expressed as learning a sparse graph of the task relationship based on prediction performance learned models. Since many natural phenomenon exhibit structures, enforcing sparsity models reveals underlying relationship. Moreover, different sparsification degrees from fully connected uncover various types like cliques, trees, lines, clusters or disconnected graphs. In this paper, we propose bilevel formulation multi-task that induces graphs, thus, revealing relationships, and an efficient method for its computation. We show empirically how induced improves interpretability their synthetic real data, without sacrificing generalization performance. Code at https://bit.ly/GraphGuidedMTL