Impact of Learning Strategies on the Quality of Bayesian Networks: An Empirical Evaluation

Abstract

We present results from a empirical evaluation of the impact of Bayesian network structure learning strategies on the learned structures. In particular, we investigate how learning algorithms with different optimality guarantees compare in terms of the structural aspects and generalisability of the produced network structures. For example, in terms of generalization to unseen testing data, we show that local search algorithms often benefit from a tight constraint on the number of parents of variables in the networks, while exact approaches tend to benefit from looser parent restrictions. Overall, we find that learning strategies with weak optimality guarantees show good performs synthetic datasets, but, compared to exact approaches, perform poorly on the more ``real-world'' datasets. The exact approaches, which guarantee to find globally optimal solutions, consistently generalize well to unseen testing data, motivating further work on increasing the robustness and scalability of such algorithmic approaches to Bayesian network structure learning.