How Deep Match to Rank Boosts CTR Prediction in E‑Commerce Recommendations

Background

In e‑commerce click optimization, the recommendation system selects items a user is most likely to click from a massive candidate pool. Typically a two‑stage pipeline is used: Matching/Candidate Generation selects a few thousand items using collaborative filtering, then Ranking predicts CTR for these candidates.

Motivation

Personalization is key for CTR prediction. This paper introduces a novel ranking model that incorporates collaborative‑filtering ideas from the matching stage into the ranking stage to better represent user‑to‑item (U2I) relevance, improving personalization and achieving strong results.

U2I Relevance Modeling

Unlike search, recommendation lacks explicit user intent; the intent is hidden in the user behavior sequence, which serves as the query. Prior works such as DIN and DIEN model user interest. Our work goes further by directly modeling U2I relevance, which measures a user’s preference strength for a target item.

U2I relevance can be modeled via collaborative filtering. I2I CF pre‑computes item similarity; U2I CF uses factorization (inner product of user and item embeddings). Recent deep models like NAIS and DNN4YouTube extend these ideas. We build two sub‑networks—User‑to‑Item and Item‑to‑Item—to capture U2I relevance.

Model Overview (DMR)

The Deep Match to Rank (DMR) model augments a standard MLP with explicit U2I cross features generated by the two sub‑networks.

User‑to‑Item Network

Inspired by factorization, we compute the inner product of user and item representations as an explicit cross feature. User representation is derived from the behavior sequence via attention‑weighted pooling, where positional encoding serves as the query. The weighted sum yields a fixed‑length vector passed through a fully‑connected layer.

Item representation is obtained via a separate embedding matrix. The inner product of user and item representations gives the U2I relevance score r.

Item‑to‑Item Network

This network computes item‑to‑item similarity to indirectly model U2I relevance, using additive attention over the behavior sequence with the target item as query. The summed attention weights provide another U2I relevance signal, which is fed to the MLP.

Training Objective

The User‑to‑Item network is trained as a large‑scale multi‑class prediction of the next item in the sequence (negative sampling). Its loss is added to the final ranking loss, encouraging larger inner‑product scores to correspond to stronger relevance.

Experiments

Offline experiments on Alibaba’s public dataset and the 1688 “Recommend for You” production dataset show consistent improvements. Online A/B testing on the 1688 platform reports a 5.5% CTR lift and a 12.8% DPV lift over the baseline DIN model, leading to full deployment.

Conclusion

The Deep Match to Rank framework jointly trains matching and ranking, and the U2I relevance modules can be easily integrated into existing CTR models as additional effective features. Future work will continue to extend the framework with new improvements.