Understanding consistency in hybrid causal structure learning

We consider causal structure learning from observational data. The main existing approaches can be classified as constraint-based, score-based and hybrid methods, where the latter combine aspects of both constraint-based and score-based approaches. Hybrid methods often apply a greedy search on a restricted search space, where the restricted space is estimated using a constraint-based method. The restriction on the search space is a fundamental principle of the hybrid methods and makes them computationally efficient. However, this can come at the cost of inconsistency or at least at the cost of a lack of consistency proofs. In this paper, we demonstrate such inconsistency in an explicit example. In spite of the lack of consistency results, many hybrid methods have empirically been shown to outperform consistent score-based methods such as greedy equivalence search (GES). We present a consistent hybrid method, called adaptively restricted GES (ARGES). It is a modification of GES, where the restriction on the search space depends on an estimated conditional independence graph and also changes adaptively depending on the current state of the algorithm. Although the adaptive modification is necessary to achieve consistency in general, our empirical results show that it has a relatively minor effect. This provides an explanation for the empirical success of (inconsistent) hybrid methods.