Research map / Label-efficient and robust learning / Recognition and long-tail
Depression Detection via Capsule Networks with Contrastive Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeCapsNet, a hierarchical capsule network with user- and post-level contrastive learning that detects depression from social media posts via PHQ-9 symptom descriptions, trained end-to-end.
Problem
Depression detection from online user posts that is both interpretable and accurate, avoiding black-box models and the two-stage pipelines that depend on costly symptom annotations.
Approach
Extract nine PHQ-9 symptom capsules from risky posts selected by Sentence-BERT similarity to symptom descriptions, route them via dynamic routing into two depression capsules, and train end-to-end with max-margin loss plus user-level and post-level supervised contrastive losses and an orthogonal attention constraint.
Key results
- Within-dataset, DeCapsNet improves F1 over the strongest baseline PHQ9plus by 6%, 9% and 5% on eRisk2018, RSDD and TRT respectively.
- Cross-dataset (RSDD to eRisk2018), DeCapsNet gains up to 8% in F1 and AUC over the second-best method.
- Ablation shows removing the user-level contrastive loss drops TRT AUC from 0.87 to 0.80.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the symptom-diversity loss Lsym (ablation) | DeCapsNet · eRisk2018, RSDD, and TRT | F1: 0.78 ± 0.07 -> 0.79 ± 0.02; 0.62 ± 0.04 -> 0.64 ± 0.01; 0.75 ± 0.01 -> 0.77 ± 0.01 | Helped |
| add user-level contrastive loss Luc (ablation) | DeCapsNet · eRisk2018, RSDD, and TRT | F1: 0.74 ± 0.02 -> 0.79 ± 0.02; 0.58 ± 0.01 -> 0.64 ± 0.01; 0.72 ± 0.00 -> 0.77 ± 0.01 | Helped |
| add post-level contrastive loss Lpc (ablation) | DeCapsNet · eRisk2018, RSDD, and TRT | F1: 0.76 ± 0.02 -> 0.79 ± 0.02; 0.59 ± 0.01 -> 0.64 ± 0.01; 0.73 ± 0.00 -> 0.77 ± 0.01 | Helped |
Limitations
- TRT dataset is not publicly released and had to be reconstructed, so results may not be exactly comparable.
- Evaluation limited to three datasets and English Reddit-style posts.
- Some baseline results are taken from prior work rather than re-run.
Experimental setup
Datasets: eRisk2018 RSDD TRT
Benchmarks: eRisk2018 RSDD TRT
Metrics: F1 AUC
Baselines: BERT-CNN Pattern (threshold) Pattern (CNN) PHQ9 (scores) PHQ9 (vectors) PHQ9plus DeCapsNet w/o Lsym DeCapsNet w/o Luc DeCapsNet w/o Lpc
Builds on: Capsule networks (Sabour et al. 2017) Supervised contrastive learning (Khosla et al. 2020) PHQ-9 based depression detection (Nguyen et al. 2022) Sentence-BERT
Closest papers by meaning
- DepMGNN: Matrixial Graph Neural Network for Video-based Automatic Depression Assessment (AAAI 2025 · similarity 0.82)
- Mental-Perceiver: Audio-Textual Multi-Modal Learning for Estimating Mental Disorders (AAAI 2025 · similarity 0.82)
- MAMA-Memeia! Multi-Aspect Multi-Agent Collaboration for Depressive Symptoms Identification in Memes (AAAI 2026 · similarity 0.80)
- MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis (NeurIPS 2025 · similarity 0.80)
- Breaking the Batch Barrier (B3) of Contrastive Learning via Smart Batch Mining (NeurIPS 2025 · similarity 0.80)
- OpticalDR: A Deep Optical Imaging Model for Privacy-Protective Depression Recognition (CVPR 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.