Boosting E‑Commerce Re‑ranking Diversity and Accuracy with Mutual‑Information

Re‑ranking rearranges product lists by considering inter‑item relationships to better satisfy user needs. Existing methods improve scoring accuracy at the expense of diversity, or boost diversity while hurting precision. To address this trade‑off, we propose PODM‑MI, a mutual‑information‑based preference‑oriented diversity model that simultaneously accounts for accuracy and diversity during re‑ranking.

Background and Current Situation

Users exhibit different decision stages (browsing, buying) with varying diversity requirements, yet current models do not directly link decision stages to diversity.

Re‑ranking must adapt to personalized user demands: high diversity when users seek varied items, high accuracy when they target a specific category.

Challenges include accurately modeling evolving user intent across multiple queries and aligning results with these dynamic intents.

PODM‑MI

The model takes a ranked list and user behavior data (clicks, cart additions) as input. First, a Preference‑Oriented Network (PON) captures user diversity preferences and candidate item diversity representations. Then, a Similarity Alignment Module (SAM) enhances consistency between user preferences and item diversity, producing a utility matrix that adaptively re‑orders results.

2.1 PON – User Preference Modeling

Historical queries and associated items provide valuable signals of user intent. Our approach incorporates click streams, cart actions, and query trajectories. Unlike static embeddings, we model user preferences with a multivariate Gaussian distribution characterized by a mean vector and diagonal covariance, capturing uncertainty and dynamic trends. Larger variances indicate broader preferences, while smaller variances reflect concentrated interests.

2.2 SAM – Optimizing Rankings with Mutual Information

After modeling preferences and item diversity, we quantify their relevance using mutual information, maximizing the shared information between user preferences and candidate items. Because direct estimation is intractable, we adopt a variational posterior estimator to derive a feasible lower bound for the mutual‑information objective.

We then learn a utility matrix via a learnable weight matrix multiplied by aligned features, and combine it with backbone scores to produce the final ranking.

2.3 Optimization Objective and Final Loss

The overall loss comprises a PRM classification loss and a mutual‑information loss, jointly guiding the model toward accurate and diverse results.

2.4 Experimental Results and Visualization

Online A/B tests on JD.com’s main search engine show that PODM‑MI increases user purchase likelihood and enhances result diversity. A 0.10% uplift in UCVR translates to substantial revenue gains.

TSNE visualizations of reduced‑dimensional query flows reveal clear stratification: divergent query streams produce more diverse ranking outcomes, while convergent streams yield more focused results, confirming the model’s ability to adapt to user intent dynamics.

Case studies illustrate that for highly diverse query histories (e.g., "Switch, Zelda, phone case, hammer, range hood"), PODM‑MI generates more varied results for a novel query, whereas for focused histories (e.g., repeated "dress" searches), the model produces concentrated, intent‑aligned rankings.

Future Directions

Incorporate finer‑grained features to better model browsing‑buying intent.

Further refine explicit influence of user intent modeling.

Explore explicit mechanisms for intent updates.