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

Jul 16, 2026

Upload by Ivy

Online social media posts offer scalable signals for early depression screening, yet common screening-stage designs often defer final decisions to a single detector. This overlooks heterogeneous user expression after screening: a monolithic classifier averages across users, diluting localized evidence and increasing errors, especially for non-self-disclosing users. WPG-MoE introduces weak-prior-guided dense mixture-of-experts on a shared LLM backbone. It learns routing using LUPI with LLM-extracted structured evidence for training-time priors, while inference relies on deployable PHQ-9 template screening plus the backbone. Experiments on Chinese and English datasets show improved performance with interpretable routing behavior.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages23

Reading Time58 min

Abstract

1 Introduction

2 The Approach

2.1 Problem Definition

WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection Xian Li 1 ,2 Yuanhe Tian2 * Yang Yang 1 Guoqing Wang 1 Yan Song3 1University of Electronic Science and Technology of China 2Zhongguancun Academy 3University of Science and Technology of China [xianli@stu.uestc.edu.cn](xianli@stu.uestc.edu.cn) [yhtian94@gmail.com](yhtian94@gmail.com) [yang.yang@uestc.edu.cn](yang.yang@uestc.edu.cn) [gqwang0420@uestc.edu.cn](gqwang0420@uestc.edu.cn) [clksong@gmail.com](clksong@gmail.com) arXiv :2607 .04350v 1 [ cs .CL] 5 Jul 2026 Abstract Online social media posts provide scalable signals for early depression screening, and recent studies mainly improve pre-classification evidence through risk-post selection, symptom grounding, and clinically informed feature construction. However, these screening-stage designs often leave final decisions to a single detector, overlooking how users heterogeneously express depressive risk after screening. A monolithic classifier must average across heterogeneous users, which may dilute localized evidence and cause misclassification, especially for non-self-disclosing users. To address this issue, we propose WPG-MoE, a weak-priorguided dense mixture-of-experts framework built on a shared large language model (LLM) backbone. WPG-MoE derives user-level weak semantic priors to softly route users to experts matched to different evidence layouts. We formulate this process as learning using privileged information (LUPI): rich LLM-extracted structured evidence guides training-time routing, while inference retains only Patient Health Questionnaire-9 (PHQ-9) template screening and the deployable backbone. Experiments on Chinese and English datasets show that WPGMoE outperforms strong baselines with interpretable routing behavior. 1 Introduction Depression affects an estimated 332 million people worldwide, yet treatment coverage and minimally adequate care remain limited (Moitra et al., 2022 ; World Health Organization, 2025) . Social media histories therefore provide scalable early-identification signals, as De Choudhury et al.(2013) showed for depression detection from naturalistic user traces (Yates et al., 2017 ; Shen et al., * Corresponding Author. Figure 1: Heterogeneous evidence patterns create a mismatch for monolithic modeling. 2017 ; Guntuku et al., 2017 ; Song et al., 2018 ; Chancellor and De Choudhury, 2020 ; Nie et al., 2020 ; Diao et al., 2023 ; Hu et al., 2024) . Recent work improves user-level social media depression detection with complete-history modeling and clinically structured evidence: multimodal fusion, user-post summarization, symptomaware temporal modeling, capsule-style aggregation (Shen et al., 2017 ; Gui et al., 2019 ; Zoganet al., 2021 ; Wang et al., 2022 ; Cai et al., 2023 ; Liu et al., 2024), Patient Health Questionnaire-9 (PHQ-9) or psychiatric-scale guidance (Nguyen et al., 2022 ; Zhang et al., 2022 ; Wang et al., 2025), and large language model (LLM)-based annotation, summarization, retrieval, or explanation for clinical evidence (Wang et al., 2024 ; Lan et al., 2025 ; Tian et al., 2024a ; Ravenda et al., 2025) . Yet final predictors across pretrained language models (PLMs), sentence-embedding pipelines, capsule models, tree classifiers, and retrieval-augmented LLM agents still make one decision after screening. The bottleneck is post-screening heterogeneity, not evidence retrieval. Depressed users reveal risk through overlapping structures: diagnoses or medication, sustained symptoms, or a few high-intensity posts amid otherwise irrelevant histories (Mendes and Caseli, 2024) . These are evidence structures, not hard clinical subtypes. Figure 1 illustrates themismatch: after screening, a flat detector compresses heterogeneous signals into one representation and boundary. This averaging can dilute localized evidence and obscure weaker non-selfdisclosure patterns, motivating dense conditional specialization

FAQs

What limitation in current screening-stage designs does WPG-MoE address?

It addresses the mismatch caused by heterogeneous post-screening evidence: final decisions are made by a single detector that averages across users, diluting localized evidence and harming accuracy for non-self-disclosing users.

How does WPG-MoE route users to experts?

WPG-MoE softly routes users to experts based on user-level weak semantic priors that match different evidence layouts, using a shared LLM backbone.

What role does LUPI (learning using privileged information) play in training and inference?

Training uses privileged, LLM-extracted structured evidence to build routing priors and evidence blocks, while inference keeps only deployable PHQ-9 template screening together with the shared backbone and history-level signals.

Jul 16, 2026

Upload by Ivy