Improving rule evaluation using multitask learning

Abstract

This paper introduces DEFT, a new multitask learning approach for rule learning algorithms. Like other multitask learning systems, the one proposed here is able to improve learning performance on a primary task through the use of a bias learnt from similar secondary tasks. What distinguishes DEFT from other approaches is its use of rule descriptions as a basis for task similarity. By translating a rule into a feature vector or "description", the performance of similarly described rules on the secondary tasks can be used to modify the evaluation of the rule for the primary task. This explicitly addresses difficulties with accurately evaluating, and therefore finding, good rules from small datasets. DEFT is implemented on top of an existing ILP system and the approach is tested on a variety of relational learning tasks. Given appropriate secondary tasks, the results show that DEFT is able to compensate for insufficient training examples.