Causally Informed Factorization Machines

Factorization machines (FMs) are a class of general predictors for sparse data. One major benefit of FMs is their ability to capture the interactions across features when making recommendations. In this paper, we note that the interactions captured by existing FMs generally represent correlations in the data and we argue that such correlations, unless informed by the true causality structures underlying the data, may be spurious and may result in unwanted bias. To tackle this challenge, we propose a Causally-Informed Factorization Machine (CIFM) model that introduces a novel causal injection mechanism. CIFM leverages a priori causal knowledge, described in the form of a causal graph, to boost the representational ability of FMs and achieve better predictions. Specifically, given a (potentially learned) causal graph which describes the causal relationships among features, CIFM distills this structural information into a pairwise causal impact matrix and guides the learning process to ensure that the learned representations capture those relationships that are consistent with the causal relationships. Extensive evaluations of CIFM, along with its integrations with NeuralFM and DeepFM, conducted with synthetic and real-world data sets, demonstrate the effectiveness of causal injection in generating better recommendations.