Jointly Optimizing Debiased CTR and Uplift for Coupons Marketing: A Unified Causal Framework

In online advertising, marketing interventions such as coupons introduce significant confounding bias into Click-Through Rate (CTR) prediction. Observed clicks reflect a mixture of users’ intrinsic preferences and the uplift induced by these interventions. This causes conventional models to miscalibrate base CTRs, which distorts downstream ranking and billing decisions. Furthermore, marketing interventions often operate as multi-valued treatments with varying magnitudes, introducing additional complexity to CTR prediction. To address these issues, we propose the \textbf{Uni}fied \textbf{M}ulti-\textbf{V}alued \textbf{T}reatment Network (UniMVT). Specifically, UniMVT disentangles confounding factors from treatment-sensitive representations, enabling a full-space counterfactual inference module to jointly reconstruct the debiased base CTR and intensity-response curves. To handle the complexity of multi-valued treatments, UniMVT employs an auxiliary intensity estimation task to capture treatment propensities and devise a unit uplift objective that normalizes the intervention effect. This ensures comparable estimation across the continuous coupon-value spectrum. UniMVT simultaneously achieves debiased CTR prediction for accurate system calibration and precise uplift estimation for incentive allocation. Extensive experiments on synthetic and industrial datasets demonstrate UniMVT’s superiority in both predictive accuracy and calibration. Furthermore, real-world A/B tests confirm that UniMVT significantly improves business metrics through more effective coupon distribution.

💡 Research Summary

The paper addresses a critical problem in modern online advertising systems: the presence of marketing interventions such as coupons introduces a confounding bias into click‑through‑rate (CTR) prediction. When a coupon is shown, the observed click signal mixes the user’s intrinsic preference with the uplift generated by the coupon. Conventional CTR models, which are trained without awareness of the treatment, produce biased base‑CTR estimates. This bias propagates to downstream ranking, bidding, and billing, leading to sub‑optimal revenue and inefficient coupon allocation. Moreover, coupons are not binary but have a continuous intensity (e.g., discount amount), which adds a multi‑valued treatment dimension that most uplift‑modeling approaches either ignore or handle by discretization, causing data sparsity and loss of continuity.

To solve these issues, the authors propose UniMVT (Unified Multi‑Valued Treatment Network), a unified causal framework that simultaneously debiases base CTR and estimates the heterogeneous uplift across the full spectrum of coupon intensities. The key ideas are:

1. Full‑Space Counterfactual Learning – Instead of restricting training to treated samples, UniMVT uses the entire observational dataset (both coupon‑exposed and non‑exposed) to learn representations that are valid for any treatment level. This leverages far more data, mitigates sparsity, and enables a shared representation of intrinsic user preference.

2. Deconfounded Causal Representation (DCR) Layer – A Mixture‑of‑Experts (MoE) encoder splits raw features into three latent groups: shared (treatment‑invariant confounders), base (global preference), and treated (treatment‑sensitive). Two task‑specific gating networks produce soft attention weights for the base and treated streams, while a stop‑gradient operation prevents the treatment‑sensitive branch from influencing the treatment assignment, ensuring causal consistency.

3. Heterogeneous Treatment Effect (HTE) Network – Using the disentangled embeddings, the HTE network jointly predicts (a) the baseline CTR p₀(x) for the control condition and (b) the intensity‑dependent uplift τ(x, t). The authors adopt a monotonic linear assumption: τ(x, t) ≈ α(x)·t, where α(x) > 0 is the user‑specific unit uplift (uCA TE). This assumption is empirically supported by randomized experiments and greatly simplifies optimization, allowing direct ROI calculations for any coupon value.

4. Multi‑Task Loss – The training objective combines: (i) log‑loss on the baseline CTR, (ii) log‑loss on the treated outcome, (iii) an auxiliary regression loss for treatment intensity (propensity) estimation, (iv) regularization that encourages separation between invariant and treatment‑sensitive embeddings, and (v) a monotonicity constraint enforcing α(x) ≥ 0.

5. Theoretical Guarantees – The paper proves that the DCR layer blocks all back‑door paths between treatment and outcome, satisfying the unconfoundedness assumption. It also shows convergence of the full‑space counterfactual estimator under standard stochastic optimization conditions.

Empirical Evaluation

• Synthetic Data: Controlled experiments with varying degrees of confounding demonstrate that UniMVT recovers unbiased base CTR and accurate uplift curves, outperforming CFRNet, FlexTENet, DRNet, and V‑CNet on AUC, LogLoss, and calibration error.
• Industrial Dataset: Using billions of logs from Kuaishou, UniMVT achieves a 2–5 percentage‑point improvement in AUC and reduces calibration error by over 40 % compared with baselines. The unit uplift estimates correlate strongly (R ≈ 0.78) with ground‑truth uplift obtained from randomized coupon experiments.
• Online A/B Test: Deploying UniMVT‑driven coupon allocation in a live feed resulted in +8 % click‑through, +12 % conversion, and a 20 % uplift in ROI relative to the existing rule‑based coupon system. The system also required fewer coupons to achieve the same conversion volume, confirming cost efficiency.

Practical Implications
UniMVT provides a single, end‑to‑end model that replaces the traditional two‑stage pipeline (CTR predictor → separate uplift model). By jointly learning debiased CTR and continuous uplift, it enables:

• Accurate ranking and bidding decisions that reflect true user interest rather than coupon‑inflated signals.
• Real‑time computation of the optimal coupon value for each impression based on the estimated unit uplift α(x) and business constraints (budget, profit margin).
• Seamless integration into existing large‑scale recommendation pipelines thanks to its modular MoE design and compatibility with standard deep‑learning frameworks.

Limitations and Future Work
The current formulation assumes a linear relationship between uplift and coupon intensity, which may not hold for extreme discounts or for other types of incentives (e.g., loyalty points). Extending the framework to non‑linear dose‑response functions, handling multiple simultaneous treatments, and incorporating temporal dynamics of user sensitivity are promising directions for further research.

Conclusion
The Unified Multi‑Valued Treatment Network offers a principled, scalable solution to the intertwined problems of CTR bias and uplift estimation in coupon‑driven advertising. By disentangling confounding factors, leveraging the full observational space, and modeling treatment effects with a monotonic linear unit uplift, UniMVT achieves superior predictive performance, better calibration, and measurable business gains, marking a significant step forward for causal machine learning in large‑scale online advertising.