Improving JD Retail Recommendation Advertising Ranking with Variational Feature Learning, User Interest Network Optimization, and Global Collaborative Modeling

JD Retail's recommendation advertising team identified three core challenges in the JD APP ad ranking pipeline: cold‑start for long‑tail users and items, limited depth of user‑interest modeling, and under‑utilization of global collaborative data.

Technical Challenges

The system must handle sparse data for new users/items, capture richer user preferences beyond immediate behavior, and leverage cross‑app interaction signals that are currently ignored.

Technical Solutions

1. Variational Feature Learning Framework (VELF) : models user and item features as distributions, applies variational inference with attribute‑conditioned priors, and optimizes a loss combining cross‑entropy and KL‑divergence.

2. User‑Interest Network Optimization :

PPNet+ introduces personalized parameter nets, gate networks, and Dice activation to better capture individual preferences.

NeNet (Needle Net) adds a residual‑style module to mitigate over‑fitting to short‑term active users.

Weighted‑MMoE incorporates multi‑task learning with task‑specific gates and attention‑weighted experts for heterogeneous ad formats.

3. Global Collaborative Information Modeling :

Pre‑training commodity embeddings on the entire JD site using graph embedding (EGES) and integrating them via Faiss‑based nearest‑neighbor tables.

Gama: a gating‑adapted wavelet multiresolution analysis model that efficiently processes long exposure sequences, extracts multi‑scale signals, and aggregates them with an interest‑gate network.

Results

The combined upgrades yielded a cumulative AUC increase of over 1% in offline experiments and a clear uplift in online revenue during the 618 promotion, with individual components contributing 0.45% (network optimizations) and 0.35% (global modeling) AUC gains.

Conclusion and Outlook

The proposed framework demonstrates that variational embedding, deep interest network enhancements, and full‑domain collaborative modeling can substantially improve CTR prediction for e‑commerce advertising, and future work will explore generative data‑driven CTR models and item‑server collaborative representations.