[2D5-GS-2] Machine learning: Statistical learning

[2D5-GS-2-01] Direct Estimation of Distributional Treatment Effects Based on f-divergence and Its Interpretability

〇Koki Yashiro¹, Koh Takeuchi¹, Hisashi Kashima¹ (1. Kyoto University)

Keywords:causal inference, distributional treatment effect

In fields such as social sciences and medicine, accurately understanding the effects that a particular intervention induces on a target is a highly significant challenge. The distributional treatment effect involves capturing changes beyond the mean by focusing on the probability distribution of latent outcomes. It can be quantified using a distance scale between probability distributions. In this study, we propose a method to directly estimate distributional treatment effects based on f-divergence without relying on the estimation of probability distributions, and we verify the interpretability of this method. In experiments, the proposed method achieved smaller estimation errors compared to a two-stage estimation involving the estimation of probability distributions.