Advances in JD E‑commerce Advertising CTR Prediction: Variational Feature Learning, User Interest Network Optimization, and Global User Collaborative Modeling

Business Background Recommendation ads are a core traffic source for JD, covering products, aggregation pages, activities, stores, videos, and live streams. The quality of ad ranking (CTR prediction) directly impacts user experience and platform revenue. The 2022 June‑18 (618) homepage redesign also upgraded the ranking models.

Technical Challenges 1) Cold‑start for long‑tail users and items. 2) Deep mining of heterogeneous user interests. 3) Insufficient utilization of global collaborative signals (exposure, clicks) across JD’s multiple apps and scenarios.

Technical Solutions

1. Variational Feature Learning Framework (VELF) VELF models user and item features as distributions rather than point estimates, using variational inference with side‑attributes as parameterized priors to obtain robust embeddings for cold‑start cases. The loss combines cross‑entropy and KL‑divergence terms:

2. User Interest Network Optimization Three modules were introduced:

PPNet+ : Extends PPNet with personalized gating, incorporating user ID, item ID, category IDs, item features, cross features, and historical click/exposure sequences. Uses Dice activation and normalization in the gate network.

NeNet : A needle‑net residual structure that mitigates over‑personalization bias by preserving non‑biased feature gradients.

Weighted‑MMoE : Multi‑gate Mixture‑of‑Experts with task‑specific gates for different ad formats, sharing a common expert pool while learning weighted contributions.

3. Global User Collaborative Information Modeling Two complementary approaches were adopted:

Global Pre‑training : Graph‑based item embeddings (EGES) are pre‑trained on JD’s whole‑site data and incorporated as side‑information during CTR model training.

Exposure‑Sequence Modeling (Gama) : A wavelet‑based multiresolution analysis (MRA) extracts hierarchical signals from long exposure sequences, followed by an interest‑gate network that adaptively weights these signals. This reduces noise and computational cost while capturing real‑time user interest.

Results and Impact The combined upgrades yielded a cumulative AUC improvement of >1% (0.45% from network optimizations, 0.35% from global modeling) and a noticeable increase in online revenue. The VELF work was published at WWW 2022, and the Gama wavelet method at SIGIR 2022.

Summary and Outlook After six months of research, JD’s advertising team deployed the variational feature learning framework, interest‑network enhancements, and global collaborative modeling across the JD app before the 618 promotion, delivering significant business gains. Future directions include generative data‑driven CTR frameworks, Item‑server bucket sequencing, and dynamic user‑server representations to further deepen full‑domain user understanding.