MindGraph

# MindGraph: Graph-Guided Multimodal Temporal Reasoning for Mental Health Detection **MindGraph** is a graph-based deep learning framework for detecting mental health conditions through the temporal and multimodal analysis of user-generated content on social media platforms. It integrates textual, acoustic, and visual modalities while modeling temporal interactions through a dynamic user mental state graph. --- ## 📘 Overview This repository contains the implementation of the paper: > **"Graph-Guided Multimodal Temporal Reasoning for Mental Health Detection on Social Media"** Mental health disorders such as depression and anxiety can often be signaled through social media activity. MindGraph is designed to: - Capture **multimodal signals** from speech, text, and facial expressions - Model **temporal progression** of mental states - Exploit **graph-guided message passing** for contextual reasoning across user states - Provide interpretable, robust predictions for early mental health detection --- ## 🧠 Key Components - **Multimodal Feature Extractor** Extracts and fuses features from textual (RoBERTa), acoustic (eGeMAPS), and visual (ResNet, VGGFace) sources. - **Mental State Graph Constructor (MSGC)** Dynamically builds inter-utterance graphs based on semantic, emotional, and speaker cues. - **Graph-Guided Temporal Reasoning Module (GTMR)** Aggregates temporal and relational evidence to predict mental health conditions at both utterance and video level. - **Cross-Level Fusion (CLF)** Integrates utterance-level and video-level embeddings to boost final prediction accuracy. --- ## 📁 Project Structure ```bash.├── data/ # Preprocessed CMU-MOSEI, DAIC-WOZ, and PHQ-8 datasets├── models/ # Core models: MindGraph, GTMR, CLF├── utils/ # Graph construction, feature alignment, metrics├── scripts/ # Training, evaluation, visualization├── configs/ # Model configs and hyperparameters├── results/ # Trained checkpoints, logs, and plots└── README.md