Metaheuristics for Score-and-Search Bayesian Network Structure Learning

摘要

Structure optimization is one of the two key components of score-and-search based Bayesian network learning. Extending previous work on ordering-based search (OBS), we present new local search methods for structure optimization which scale to upwards of a thousand variables. We analyze different aspects of local search with respect to OBS that guided us in the construction of our methods. Our improvements include an efficient traversal method for a larger neighbourhood and the usage of more complex metaheuristics (iterated local search and memetic algorithm). We compared our methods against others using test instances generated from real data, and they consistently outperformed the state of the art by a significant margin.