WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection

WPG-MoE utilizes a shared LLM backbone and a weak-prior-guided routing mechanism to address user heterogeneity in social media depression detection. The framework employs Learning Using Privileged Information (LUPI) to map users to specific experts based on semantic priors, preventing localized evidence dilution inherent in monolithic classifiers.