| Dynamic Predicate Construction for Learning Relational Concepts |
| Michael Chiang, University
of British Columbia, Canada |
| The aim of this work is to enrich the search space of relational rule learning methods by dynamic feature construction, with predictive performance as the objective function. Under current methodology, a subgroup of individuals is deemed distinguishable if it is possible to describe it with a rule, which implies a priori availability of the relevant features in the learning vocabulary. When this is not the case, overly general rules result along with loss in predictive performance. Conversely, large features lead to overly complex models, which tend to overfit data. We attempt to bridge this generality gap by discovering interesting subgroups of individuals (via a dynamic programming algorithm) from which new predicates are formed, without requiring predefined features. We demonstrate that our method generates intermediate hypotheses that achieves predictive gains as well as effective leveraging of overfitting. Evaluations are carried out synthetic as well as real-world datasets. |