Offline Long-Term Causal Effect Estimation with Short-Term Experimental Data for Recommendation Systems

In recommendation systems, evaluating a recommendation policy typically involves causal inference through either online A/B testing or offline estimation of historical data. Time and cost constraints lead to policy evaluation based on short-term rather than long-term metrics - yet long-term metrics like lifetime-value (LTV) is the true outcome of interest. Real-world marketing confronts two core issues: weak short-term effects and missing new policy records in historical data. These two challenges preclude reliable long-term estimation and harm platform decisions. In this paper, we propose a method, Short-Term Amplified Estimator for Long-term effects (STEAL), that overcomes the two challenges above in long-term causal effect estimation. STEAL addresses the first challenge by learning treatment-control patterns from short-term data and matching analogous samples in long-term data. To address the second challenge, STEAL divide short-term and direct effects to enable more direct and accurate estimation of long-term causal effects. Moreover, we introduce an interactive attention mechanism into STEAL to better model the interactions between the decomposed components and user features, making STEAL adaptable to high-dimensional data commonly encountered in real-world online marketing platforms. Extensive experiments on synthetic data, semi-synthetic data, and real-world data from a large-scale e-commerce platform demonstrate the effectiveness of STEAL.