Jiandong Ding (丁建栋)

Principal algorithm expert at Huawei Technologies.

I am a Principal Algorithm Expert at Huawei Technologies, bridging the gap between theoretical algorithms and industrial-scale systems.

My research philosophy is summarized as “From Biological Sequences to User Behaviors”. I apply deep representation learning to decipher underlying patterns in data—from genomic sequences in my early career to billion-scale user behavior logs in commercial recommendation systems today.

Currently, I focus on building Next-Generation Recommender Systems driven by Generative AI, with specific interests in:

  • Generative RecSys: LLM-driven Recommendation, Sequential Modeling, and User Representation.
  • Trustworthy AI: Causal Inference, Unbiased Learning, and Fairness in Ranking.
  • System Efficiency: Edge-Cloud Collaboration, Retrieval Architecture, and Model Compression.

🔥 News

  • [Feb 2026] 🎉 Paper “Mitigating Popularity Bias in Recommendation” accepted by ACM TOIS.
  • [Jan 2026] 🚀 Paper “Hierarchical and Preference-Aware Generative Recommendations” accepted by TheWebConf (WWW) 2026.
  • [Nov 2025] Two papers accepted! “RPE4Rec” by WSDM 2026, and “Invariant Feature Learning” by AAAI 2026.

🎯 Research Highlights

My current research focuses on three core pillars: Generative & Trustworthy RecSys, Extreme Efficiency, and LLM Agents.

1. Next-Gen Recommendation: Generative & Trustworthy

WWW Framework
[WWW 2026] Generative Recs: Hierarchical & Preference-Aware Problem Existing "flat-sequence" generative models overlook the hierarchical structure of user sessions and introduce noise in long histories.

Breakthrough We proposed HPGR, a two-stage generative framework. It combines structure-aware pre-training with preference-guided sparse attention to capture the true hierarchy of user interests, achieving SOTA performance.
AAAI Framework
[AAAI 2026] Causal Inference for Watch-time Prediction Problem In short-video feeds, "duration biases" (longer videos naturally get more watch time) mislead algorithms.

Breakthrough We introduced Invariant Feature Learning based on counterfactual inference to uncover the user's true willingness to watch, independent of video length.

2. Extreme Efficiency at Scale

WSDM Framework
[WSDM 2026] RPE4Rec: High-Efficiency Dynamic Retrieval Problem Advanced Transformers are often too slow for real-time retrieval on billion-scale items.

Breakthrough We designed a novel Relative Position Encoding (RPE) mechanism specifically for dynamic node retrieval. This architecture significantly reduces inference latency while capturing complex sequential patterns.
KDD Framework
[KDD 2024] Low-Rank Compression for CTR Prediction Breakthrough A unified framework to compress massive CTR models using low-rank factorization, enabling high-performance ranking on resource-constrained devices (e.g., mobile phones).

3. LLM Agents & Data Intelligence

ICSOC Framework
[ICSOC 2025] NL2SQL Benchmark for Business Intelligence Intelligence Layer Evaluating how Large Language Models (LLMs) act as Data Agents to translate natural language into complex SQL queries, enabling automated business decision-making.
DPSSL Framework
[NeurIPS 2021] DP-SSL: Robust Semi-supervised Learning Problem Deep learning performance heavily relies on massive labeled data, which is expensive to obtain.

Breakthrough We introduced a Data Programming (DP) scheme that automatically generates probabilistic labels for unlabeled data, achieving SOTA performance with minimal supervision (only 40 labeled samples).

📫 Get in Touch

I am deeply committed to bridging the gap between academia and industry. Having led numerous research initiatives at Huawei CBG and Alibaba DAMO, I am always open to:

  • Academic Partnerships: Collaborative research & grant applications.
  • Professional Events: Industry summits and tech forums.

For collaboration inquiries: jdding [at] fudan.edu.cn