Research On Constraint-based Algorithms For Bayesian Network Structure Learning

This paper describes the status quo of Bayesian Network structure learning, introduces the origins and developments of Bayesian Network and the main features of Bayesian Networks, highlights analysis the Bayesian Network Structure learning mechanism, discusses the merits of various algorithms, and mainly studies the constraint-based Bayesian Network structure learning algorithms. High time complexity and the appearance of inconsistency during the learning process will be two major problems. The main contents are as follows:Model causal structure learning algorithms try to find a directed acyclic graph, which can represent the causal relationships among variables perfectly. Existing such constraint-based algorithms implement conditional independence tests to construct the graph structure by comparing pairs of nodes independently. In this paper, we introduce a two-phase method for learning equivalence class of Bayesian Network. First, it learns a skeleton of the Bayesian Network by conditional independence tests. In this way, it reduces the number of tests compared with other existing algorithms, and decreases the running time drastically. The second phase of our method orients edges that exist in all Bayesian Network equivalence classes. Our approach is tested on the alarm network; the experimental results show that our method outperforms other algorithms.When limiting the data to be non-interventional, it is very hard to identify the exact directed acyclic graph. Hence, typical constraint-based algorithms usually orient partial edges which form V-structures. With the consideration of noise, locality of conditional tests and statistical errors, some conflicting structures may be returned. However, such problems are not fully resolved by these algorithms. In the paper, an approach to control these conflicts is proposed. The theoretical results show that our algorithm is correct. The effectiveness of the proposed algorithm is demonstrated by the comparison with state-of-the-art algorithms through numerical experiments.