[5M2-GS-2c] Machine learning

12:15 PM - 12:30 PM JST(3:15 AM - 3:30 AM UTC)

[5M2-GS-2c-02] Identification of KPI Variation Factors via Causal Contribution Decomposition using Observational Data

〇Atomu Matsuda¹, yusaku imai¹ (1. GROWTH DATA Inc)

Keywords:

Causal Inference,simultaneous intervention,Causal Discovery

In modern business analytics, identifying the root causes of Key Performance Indicator (KPI) variations is a critical task. However, observed KPI changes are often the result of ``Compound Interventions'' where multiple factors shift simultaneously. Conventional causal inference methods, which typically focus on the average treatment effect of a single intervention, fail to decompose these total variations consistently. To address this, we propose a framework that decomposes KPI variation into the causal contribution of each factor (including intermediate variables) based solely on observational data. We define the ``Distributional-Shift Weighted Effect'' as the estimand, which integrates the distribution shift of factors with their causal response curves identified via g-computation. Experiments using realistic synthetic data and ground truth derived via Leave-One-Out (LOO) toggling demonstrate that our method achieves substantially higher F1@10 and nDCG@10 than correlation, OLS, Lasso, and permutation importance, while remaining fully observation-only.