Learning Ground CP-Logic Theories by Leveraging Bayesian Network Learning Techniques

摘要: Causal relations are present in many application domains. Probabilistic Logic (CP-logic) is a probabilistic modeling language that especially designed to express such relations. This paper investigates the learning of CP-logic theories (CP-theories) from training data. Its first contribution SEM-CP-logic, an algorithm learns CP-theories by leveraging Bayesian network (BN) techniques. SEM-CP-logic based on transformation between and BNs. That is, method applies BN techniques learn CP-theory form equivalent BN. To this end, certain modifications required parameter structure search, most important one being refinement operator used search must guarantee constructed BNs represent valid CP-theories. The paper's second theoretical experimental comparison learning. We show simple can be represented with consisting noisy-OR nodes, while more complex require close fully connected networks (unless additional unobserved nodes introduced network). Experiments controlled artificial domain latter cases requires fewer data than also apply medical context HIV research, it compete state-of-the-art methods domain.