How JD’s Explore & Exploit Module Tackles Position and Popularity Bias in Search Ranking

Background

The Explore & Exploit (EE) module in JD’s search system aims to alleviate the Matthew effect caused by data bias, improving product diversity and exploration efficiency while maintaining search efficiency.

EE Scenario Iteration Loop

Each step—from core positioning to online metrics, offline evaluation, and model iteration—must be upgraded with EE‑specific considerations.

Model Debias Iteration

Problem Background

EE seeks to explore more efficient long‑tail items, but three major biases hinder fair exposure:

Position‑bias : User behavior is heavily influenced by item position, amplifying head‑item advantages.

Popularity‑bias : Items with higher historical sales or reviews dominate when multiple candidates have similar relevance.

Exposure‑bias : Only a small subset of items is shown per query, creating a training‑inference distribution gap.

Current EE ranking models address position and popularity bias through bias‑net modeling, achieving modest gains, while exposure bias remains an open problem.

Position‑bias Modeling

The EE model adopts a two‑stage position‑debias scheme:

Pos as feature : During training, position is used as an input feature; during inference, all items share a forced position value to neutralize positional effects.

Multi‑position prediction : A multi‑channel output predicts logits for each possible position, masking non‑actual positions during training and selecting items greedily during inference.

Pos tower : A dedicated position‑bias network is added to the DNN+SVGP scoring model, with separate training and prediction phases to remove positional influence.

Risks include choosing an appropriate forced position range and ensuring the positional feature’s impact is adequately reflected in the final logits.

Personalized Position‑bias Modeling

Users differ in sensitivity to position; a personalized bias‑net incorporates static user profiles and dynamic behavior sequences to compute user‑specific position preferences, better restoring true content preference.

SVGP Introduction

Sparse Variational Gaussian Process (SVGP) handles large‑scale covariance computation by learning a set of inducing points, enabling efficient mean and variance estimation for unseen samples.

Representation Fusion & Logit Fusion

Two fusion strategies are explored:

Representation fusion : Concatenating or summing position and content embeddings before the final layer.

Logit fusion : Combining content and position logits via addition or multiplication, directly reflecting positional gain.

Online AB Metrics

With overall traffic efficiency unchanged, EE core metrics improved markedly: exploration liquidity (+1.35%) and exploration success rate (+0.74%). After causal debiasing, exploration success rose further (+0.82%) while maintaining efficiency.

Offline Evaluation System

Custom offline metrics assess efficiency, long‑tail exploration intensity, and uncertainty estimation, bridging the gap between online success and offline AUC‑type scores.

Conclusion

EE’s bias‑focused debiasing—addressing position, popularity, and exposure biases—significantly boosts search diversity and long‑tail item exposure. Ongoing work includes richer exploration signals, long‑term value integration, and candidate set optimization.