An Improved Incremental Structure Learning Algorithm for Bayesian Networks

Abstract

As traditional batch Bayesian structure algorithm may not meet the the new data stream. In such environment, it is important to learn Bayesian structure by modifying the current structure incrementally. In this paper, a new incremental learning algorithm for Bayesian network structures (IBN) is proposed. IBN learning is decomposed into three phases: improved batch Bayesian structure learning, structure matching judgment and structure revision. In the improved batch learning procedure, improved immune operator, improved self-adaptive crossover operator and mutation operator are embedded to simplified hill climbing algorithm, so that obtain a best structure which has better performance in accuracy and efficiency, at the same time a function is constructed to monitor new data stream. In the structure matching judgment procedure, a new evaluation criterion “IBIC” (implicit Bayesian information criterion) is proposed based on BIC and implicit method. In the structure refinement procedure, the mutation and climbing operators are used to correct existing structure automatically. Intensive experiments were conducted to evaluate our algorithm, compare with incremental Max-Min parents and Children (iMMPC) and improved hill climbing search(iHCS). Experiments results shown that accuracy and efficiency of our algorithm are better than iMMPC and iHCS. Besides, this paper provided a new way for incremental Bayesian structure learning in the future.