Research mainline

My work has moved from structured biological data, through robust learning and analytics, to recommendation and agent systems used at industrial scale.

Zero-observation user reactivation, generative recommendation, dynamic retrieval, semantic-ID diagnostics, and agent skill retrieval.

BI/NL2SQL evaluation, LLM serving, CTR model efficiency, and large-scale recommendation infrastructure.

Continual graph learning, neural topic modeling, weak supervision, robust learning, and live-streaming field experiments.

miRNA target prediction, genome-scale sequence signals, and early structured data mining.

Selected papers

Representative papers that mark the main research line across recommendation, LLM agents, and data mining.

Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations
WWW 2026 Recommender Systems

Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations

Hierarchical and preference-aware generative recommendations move beyond flat user histories and expose structured intent over time.

Invariant feature learning for counterfactual watch-time prediction in video recommendation
AAAI 2026 Recommender Systems

Invariant feature learning for counterfactual watch-time prediction in video recommendation

Invariant feature learning separates duration effects from true user preference in short-video recommendation.

RPE4Rec: Enhancing Dynamic Node Retrieval with Efficient Relative Position Encoding for Recommendation Systems
WSDM 2026 Recommender Systems

RPE4Rec: Enhancing Dynamic Node Retrieval with Efficient Relative Position Encoding for Recommendation Systems

Efficient relative position encoding for dynamic node retrieval in recommendation systems.

BIS: NL2SQL Service Evaluation Benchmark for Business Intelligence Scenarios
LNCS 2025 LLM Agents

BIS: NL2SQL Service Evaluation Benchmark for Business Intelligence Scenarios

A benchmark for evaluating service-oriented NL2SQL systems in business intelligence scenarios.

Unified Low-rank Compression Framework for Click-through Rate Prediction
KDD 2024 Recommender Systems

Unified Low-rank Compression Framework for Click-through Rate Prediction

A unified compression framework for large CTR models and resource-constrained deployment.

Active projects

Current research directions are intentionally described at a high level; mature public outputs remain in Full publications and Patents.

Recommendation

Dormant-user recommendation

Studying recommendation for returning users after long inactivity gaps, where pre-gap history exists but its current reliability must be recalibrated.

Search/recommendation

Search and recommendation interaction

Studying how search behavior and recommendation exposure interact, with emphasis on attribution boundaries rather than headline metric movement.

Semantic IDs

Semantic-ID auditing

Auditing semantic-ID mappings used by generative recommendation systems before expensive downstream training and deployment decisions.

Agent skills

Agent skill retrieval and governance

Studying retrieval and governance problems when agent skills become reusable software assets in larger enterprise skill libraries.

Collaboration and exchange

Open to university collaboration, research exchange, invited talks, and focused discussions around recommendation, agent systems, and data intelligence.

Research

Academic collaboration

University collaboration, joint research, resource building, benchmark design, and student or lab exchange.

Industry

Industrial research

Recommendation architecture, agent evaluation, retrieval systems, and data-intelligence problems at production scale.

Talks

Invited talks

Research talks and professional events on recommender systems, LLM agents, and data mining.