JD's Two Papers Accepted at IJCAI2020 and SIGIR2020: Hierarchical Reinforcement Learning for Multi‑Goal Recommendation and Attention‑Based pCVR Prediction

Recently, JD had two technical papers accepted as full papers at top conferences IJCAI2020 and SIGIR2020, underscoring its strong AI research strength. IJCAI is the flagship AI conference with a 12.6% acceptance rate in 2020, while SIGIR is the leading venue for information retrieval and data mining.

JD’s Vice President and President of the Business Promotion Division, Dr. Yan Weipeng, emphasized that AI technology is a core driver of JD’s rapid growth and will continue to boost efficiency across industries and society.

The two accepted papers originate from JD’s Business Promotion Division, which leverages AI, big‑data analytics, and full‑network resource integration to provide intelligent marketing services for merchants, suppliers, and partners. Over the years the division has produced dozens of papers at top venues and cultivated many AI experts.

Paper 1: Hierarchical Reinforcement Learning for Multi‑Goal Recommendation (MaHRL) – SIGIR2020

Traditional e‑commerce recommendation systems focus mainly on click‑through rates, neglecting conversion. JD proposes MaHRL, a hierarchical reinforcement‑learning framework that jointly optimizes clicks and conversions. A low‑level agent learns short‑term interests (exposure/click), while a high‑level agent captures long‑term interests (sparse click‑conversion signals) and feeds them as features and rewards to the low‑level agent, guiding it toward conversion‑oriented ranking.

MaHRL has been deployed in JD’s advertising trigger ranking, delivering significant improvements in ad revenue and conversion‑rate metrics, thereby strengthening the ecosystem for users, media, and advertisers.

Paper 2: Attention‑Based Model for Conversion Rate Prediction with Delayed Feedback (pCVR) – IJCAI2020

Accurate pCVR estimation is crucial for JD’s smart bidding (e.g., target CPA, eCPM). The paper addresses two challenges: extreme sparsity and non‑linearity of user behavior data, and variable time delays between click and conversion. The proposed solution is a two‑tower deep learning architecture that extracts item embeddings from display/click data, employs an attention mechanism to model user purchase intent, and incorporates a time‑delay model trained jointly with the conversion model to correct label bias caused by delayed feedback.

In production, the enhanced pCVR model markedly improves prediction accuracy and overall advertising performance.

“Every technological revolution brings disruptive changes to advertising,” said Yan Weipeng, adding that AI’s rapid development has already permeated JD’s entire advertising workflow and that the Business Promotion Division will continue to mine data value and explore AI applications to build a data‑driven marketing ecosystem.