A combined approach to multi-label multi-task learning

Delaram Motamedvaziri, Venkatesh Saligrama, David A. Castañón

Abstract: In this paper, we present a method for jointly learning r >; 1 similar classification tasks. We consider a set of classification tasks whose relevant features may have some overlap. This potential overlap encourages the idea of learning tasks simultaneously. Our method is based on the idea of using two regularizers which control the underlying structure of the model from completely unrelated tasks to practically the same tasks. We show that this problem is equivalent to a convex optimization problem. Our results on simulated and real data sets demonstrate that our proposed method dramatically improves the performance on partially related tasks in comparison to independently learning the tasks or other multi-task approaches.