Professor Sungkyu Park's research paper, titled "Mitigating Spurious Correlations in Mental Health Analysis: A Contrastive Learning Approach for Cross-Domain Robustness," was presented at the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2026), held in Jeju, Korea, on August 9–13. KDD is one of the premier international conferences in data mining and artificial intelligence. The study addresses a critical reliability problem in AI-based mental health analysis: language models that detect mental disorders from social media text often rely on spurious correlations - superficial artifacts of the training data rather than genuine clinical signals - causing their accuracy to collapse when applied to new domains such as counseling conversations. The proposed framework combines artifact-masked contrastive learning with dynamic hard-negative sampling, achieving the highest accuracy on out-of-distribution conversational benchmarks among all compared methods while matching the strongest baselines in-domain and outperforming much larger 7B-scale generative models at a fraction of the size.

The findings suggest a practical path toward deploying trustworthy mental health monitoring tools in social media and conversational AI systems. 

You can find the full paper in the following link:https://dl.acm.org/doi/abs/10.1145/3770855.3818907