Dr. Yushun Dong Received Awards (SIGSPATIAL, ICDM) and Has Papers Published
(AAAI ×2, WSDM, SIGKDD)
Dr. Yushun Dong has received multiple major research awards and published several papers at premier international conferences in data mining, artificial intelligence, and geospatial analytics.
Two Awards at Premier Data Mining & Geospatial Venues
SIGSPATIAL 2025: Best Short Paper Award for TyphoFormer
Dr. Yushun Dong’s research team won the Best Short Paper Award at ACM SIGSPATIAL 2025, one of the top-tier conferences in geospatial analysis. The awarded paper,TyphoFormer: Language-Augmented Transformer for Accurate Typhoon Track Forecasting, introduces a multimodal forecasting framework that integrates meteorological time-series with natural language prompts generated from large language models. This paper establishes a new paradigm for language-enhanced geospatial forecasting, with strong implications for climate resilience and disaster preparedness.
ICDM 2025 BlueSky Track: Second Prize for Research Vision on Explainability & Extractability
At the IEEE ICDM 2025 BlueSky Track, Dr. Dong’s team received the Second Prize for the visionary paper Navigating Between Explainability and Extractability in Machine Learning as a Service. This work identifies a fundamental tension in modern ML systems: regulatory pressure demands more transparency, yet explainability features can increase vulnerability to model extraction attacks.
Four Publications at AAAI 2026, WSDM 2026, and KDD 2026
WSDM 2026: MolEdit — Knowledge Editing for Multimodal Molecule Language Models
Dr. Dong’s team introduced MolEdit, a novel framework enabling fine-grained knowledge editing for multimodal molecule language models. MolEdit supports targeted correction and refinement of chemical knowledge representations without retraining from scratch, offering a powerful tool for accelerating molecular discovery pipelines in drug design, materials science, and chemistry-driven AI. This work positions multimodal LLMs as reliable scientific assistants capable of dynamic self-correction and domain adaptation.
AAAI 2026: Query-Efficient Domain Knowledge Stealing Against Large Language Models
This paper investigates a novel class of query-efficient domain knowledge extraction attacks that target large language models. The team demonstrates that attackers can efficiently approximate domain-specific knowledge from black-box LLM APIs under tight query budgets. The work not only provides the first systematic evaluation of such attacks but also highlights urgent implications for intellectual property protection in LLM deployment—impacting cloud AI services, enterprise LLMs, and high-value scientific models.
AAAI 2026: Targeted Pathway Inference for Biological Knowledge Bases via Graph Learning and Explanation
This interdisciplinary research combines graph machine learning with explainable AI to infer biologically meaningful pathways in large-scale biomedical knowledge bases. The proposed model delivers interpretable, high-fidelity pathway predictions, enabling domain experts to examine molecular mechanisms with both computational rigor and scientific transparency.
The work expands the frontier of explainable graph learning for computational biology and precision medicine.
KDD 2026: Certified Defense on the Fairness of Graph Neural Networks
This paper introduces the first certified defense framework for fairness based on graph neural networks. The framework provides provable guarantees on fairness metrics under adversarial perturbations, establishing a mathematical foundation for fair graph learning systems. This contribution represents a critical advance for high-stakes applications such as finance and healthcare, where model fairness needs to withstand adversarial manipulation.