The use of causal inference with structural models in industry

Author

Isozaki, Takashi

Abstract

This article describes the use of probabilistic structural causal models in industry. With the accumulation of observational data and the development of inference methods in the models, it has become possible to infer causal relationships to a considerable extent if excellent inference algorithms are available. In industry, the key objective is not necessarily to infer causal relationships per se, but rather, to ensure that the inference of causal relationships leads to knowledge of the factors that have a direct or fundamental impact on the items to be improved, thereby leading to more efficient interventions. In this task, it is more important to discriminate between pseudo correlation and causation, unlike in prediction tasks, and it is also useful to distinguish between direct and indirect factors. This article first briefly reviews the theory, methods, and algorithms of causal inference that provide the foundations for its utilization, and then presents our approach to addressing a longstanding problem in this area. We then provide some discussion on the modeling and analytics of what we have been paying attention to when using the technologies in actual data analysis. After that, we introduce several publicly available examples of analyses using causal inference at both Sony and outside companies (e.g., companies involved in semiconductors, automobiles, plating and other manufacturing, product marketing, and analysis of machine learning in the music field). Our examples demonstrate that causal inference is effective in a broad range of business settings.