A taxonomy of weight learning methods for statistical relational learning

Abstract Statistical relational learning (SRL) frameworks are effective at defining probabilistic models over complex data. They often use weighted first-order logical rules where the weights of govern interactions and usually learned from Existing weight approaches typically attempt to learn a set that maximizes some function data likelihood; however, this does not always translate optimal performance on desired domain metric, such as accuracy or F1 score. In paper, we introduce taxonomy search-based for SRL directly optimize chosen metric. To effectively apply these approaches, novel projection, referred scaled space (SS), is an accurate representation true space. We show SS removes redundancies in captures semantic distance between possible configurations. order improve efficiency search, also approximation which simplifies process sampling demonstrate two state-of-the-art frameworks: Markov logic networks soft logic. perform empirical evaluation five real-world datasets evaluate them each different metrics. compare against four other approaches. Our experimental results our proposed outperform likelihood-based yield up 10% improvement across variety Further, extensive measure robustness approach initializations hyperparameters. The indicate both robust.