A Posterior-Based Method for Markov Logic Networks Parameters Learning

The past few years have witnessed a significant development in statistical relational learning (SRL). Markov logic network (MLN), one of SRL methods, is a first-order knowledge base with a weight attached to each formula, and could be viewed as a template for ground Markov logic networks. In this paper, a posterior-based parameters learning approach for Markov logic networks, maximum pseudo-posterior estimation is proposed. Mean Gaussian distribution is used as prior of each weight, likelihood is replaced by pseudo-likelihood, and the pseudo-posterior distribution is maximized to learn the weights. Experiments show maximum pseudo-posterior estimation could learn MLNs model effectively, which performs inference better compared to maximum pseudo-likelihood estimation