Reviews: Inferring Generative Model Structure with Static Analysis

The authors consider the setting where we wish to train a discriminative model using labels that generated using so-called heuristic functions, which in turn make use of primitive features. In order to generate a single label that combines multiple heuristics, the authors learn a probabilistic model (represented as a factor graph). This probabilistic model incorporates two types of structure. The method performs static analysis on the heuristic functions to identify cases where multiple functions make use of the same primitives. In order to capture correlations between primitives, the model learns pairwise similarities.