Alibaba's DANet boosts e-commerce CVR by 3.63% with discount awareness

Conversion rate (CVR) prediction is a cornerstone of e‑commerce recommendation systems, yet most models overlook a critical factor: item discount rates. Discounts heavily influence both pricing strategies and user purchasing behavior, but existing CVR models treat discounts as just another feature. Researchers from Alibaba now propose the Discount-Aware Network (DANet), which explicitly models the relationship between discount rates and post‑click conversion probability. DANet comprises three specialized components: (1) a time‑frequency transformation module that applies Fourier transforms to capture long‑term discount trends from the frequency spectrum of discount histories; (2) a distribution de‑bias module that corrects biases arising from varied purchase combinations, promotional activities, and periodic deviations across different promotion periods; and (3) a supervised regression auxiliary task that creates explicit discount labels to improve value accuracy and representation. Together, these components address data sparsity and sample selection bias while leveraging discount dynamics often ignored in prior work.

Experimental results on real e‑commerce datasets demonstrate DANet’s superiority. Offline Area Under the Curve (AUC) improved by 1.61%, a statistically significant gain for large‑scale systems. An online A/B test on Alibaba’s Tmall app showed even more compelling results: pCVR (post‑click conversion rate) increased by 3.63% and Gross Merchandise Volume (GMV) rose by 2.23%. These gains confirm that discount‑aware modeling directly translates to revenue growth. DANet has been successfully deployed on Tmall, serving millions of users daily. The team has released the code, enabling other platforms to adopt similar techniques. This work highlights a promising direction for recommendation systems: incorporating economic signals like discounts to better predict and influence user behavior.