Title: A Unified Causal Approach for Simultaneous Debiased CTR and Uplift Optimization in Coupon Marketing

Abstract: The deployment of marketing interventions, such as coupons, in online advertising introduces substantial confounding bias into Click-Through Rate (CTR) prediction models. Because observed clicks are a composite of users’ inherent preferences and the uplift generated by these interventions, standard models often fail to accurately calibrate base CTRs. This miscalibration subsequently distorts critical downstream processes, including ranking and billing. Additionally, marketing campaigns frequently function as multi-valued treatments with diverse magnitudes, adding another layer of complexity to the prediction task. To resolve these challenges, we introduce the Unified Multi-Valued Treatment Network (UniMVT). This architecture isolates confounding variables from treatment-sensitive representations, facilitating a comprehensive counterfactual inference module that simultaneously reconstructs debiased base CTRs and intensity-response curves. To manage the intricacies of multi-valued treatments, UniMVT utilizes an auxiliary task for estimating intensity to capture treatment propensities, alongside a unit uplift objective designed to normalize the intervention effect. This approach guarantees consistent estimation across the continuous spectrum of coupon values. By concurrently delivering debiased CTR predictions for system calibration and precise uplift estimates for incentive allocation, UniMVT offers a dual benefit. Extensive evaluations on both synthetic and industrial datasets highlight UniMVT’s superior performance in predictive accuracy and calibration. Moreover, real-world A/B testing validates that UniMVT enhances business outcomes through more efficient coupon distribution.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC