Who Learns Better Bayesian Network Structures: Accuracy and Speed of Structure Learning Algorithms

TL;DR

This paper compares the speed and accuracy of constraint-based, score-based, and hybrid algorithms for learning Bayesian network structures, revealing that algorithm performance is heavily influenced by the statistical criteria used.

Contribution

It provides a comprehensive evaluation of the three classes of structure learning algorithms, highlighting the impact of statistical criteria on their performance beyond algorithm design.

Findings

Constraint-based algorithms are less accurate than score-based ones.

Constraint-based algorithms are not faster than score-based algorithms at large sample sizes.

Hybrid algorithms do not outperform constraint-based algorithms in speed or accuracy.

Abstract

Three classes of algorithms to learn the structure of Bayesian networks from data are common in the literature: constraint-based algorithms, which use conditional independence tests to learn the dependence structure of the data; score-based algorithms, which use goodness-of-fit scores as objective functions to maximise; and hybrid algorithms that combine both approaches. Constraint-based and score-based algorithms have been shown to learn the same structures when conditional independence and goodness of fit are both assessed using entropy and the topological ordering of the network is known (Cowell, 2001). In this paper, we investigate how these three classes of algorithms perform outside the assumptions above in terms of speed and accuracy of network reconstruction for both discrete and Gaussian Bayesian networks. We approach this question by recognising that structure learning is defined by the combination of a statistical criterion and an algorithm that determines how the criterion is applied to the data. Removing the confounding effect of different choices for the statistical criterion, we find using both simulated and real-world complex data that constraint-based algorithms are often less accurate than score-based algorithms, but are seldom faster (even at large sample sizes); and that hybrid algorithms are neither faster nor more accurate than constraint-based algorithms. This suggests that commonly held beliefs on structure learning in the literature are strongly influenced by the choice of particular statistical criteria rather than just by the properties of the algorithms themselves.