Enterprise AI Analysis

TSCAN: A Novel Two-Stage, Context-Aware Uplift Model

TSCAN (Two-Stage Context-Aware Uplift Network) is a novel framework designed to overcome the limitations of single-stage ITE models. It comprises two sub-models: CAN-U (Context-Aware Uplift Network - Unbiased) and CAN-D (Context-Aware Uplift Network - Direct). Both leverage a shared backbone architecture, including a Feature Encoder, Context-Aware Attention Layer, and Isotonic Output Layer. The key innovation is how these stages are used to first mitigate bias and then optimize predictive performance, without the constraints of bias-mitigation techniques affecting the final prediction.

Decoupling Bias Mitigation from Prediction

TSCAN employs a crucial two-stage training strategy:

• Stage 1 (CAN-U): This model is trained with Integral Probability Metric (IPM) and propensity score regularization to generate high-quality pseudo-uplift labels. This stage focuses on mitigating selection bias and ensuring causal robustness.
• Stage 2 (CAN-D): Leveraging the pseudo-uplift labels from Stage 1, CAN-D performs supervised uplift learning. Crucially, it removes the regularizations used in Stage 1, preventing information loss and performance degradation. It directly models uplift effects using an isotonic output layer, allowing adaptive correction of estimation errors through factual outcome reinforcement. This separation allows TSCAN to enjoy the benefits of bias correction while avoiding its pitfalls in final prediction.

Dynamic Contextual Adaptation for Precision

A core innovation in TSCAN is the Context-Aware Attention Layer. This mechanism dynamically fuses embeddings of merchant features, treatment variables, and external contextual factors (e.g., time of day, district type, supply-demand ratio). This allows TSCAN to capture context-dependent heterogeneity in treatment effects. For example, a promotional offer might be highly effective during off-peak hours but have minimal impact during peak demand due to market saturation. This layer enables the model to adapt its predictions to varying operational environments, ensuring more precise and relevant business diagnoses.

Rigorous Benchmarking & Ablation Studies

TSCAN was rigorously evaluated on two large-scale real-world datasets from Taobao Shangou (Ele.me):

• Eleshop-1M (1M samples): Continuous treatment (merchant rating), outcome (order count).
• Shop Activities (700k samples): Binary treatment (marketing activity participation), outcome (order count).

TSCAN consistently outperformed seven state-of-the-art baselines across all metrics (QINI, AUUC, CAUUC, CQINI). The ablation studies confirmed the critical contributions of the two-stage training, Context-Aware Attention Layer, and Isotonic Output Layer to TSCAN's superior performance and its ability to capture context-dependent effects.

0.76% Order Increase in Live A/B Test

TSCAN was deployed in a live A/B test on one of China's largest online food ordering platforms for merchant diagnosis. In comparison to the previously deployed BART model, TSCAN achieved:

• A 0.76% increase in merchant order volume.
• Significant improvements in both Context-Aware AUUC (+0.0411) and overall AUUC (+0.0349).

The model's adaptive behavior was evident: during peak demand in business districts, it down-weighted discount suggestions while up-weighting exposure-boosting ones; conversely, during low demand in residential areas, it recommended targeted discounts. This validates its practical utility and impact for delivering context-specific, personalized business recommendations.