Multi-Agent Causal Reasoning for Suicide Ideation Detection Through Online Conversations
New AI framework uses counterfactual reasoning and bias mitigation to detect suicide risk in online conversations.
A research team led by Jun Li has introduced a novel Multi-Agent Causal Reasoning (MACR) framework designed to improve suicide risk detection in online conversations. The system addresses two critical limitations of current methods: their reliance on narrow, predefined rules for capturing user interactions, and their failure to account for hidden social influences like user conformity and suicide copycat behavior that can significantly affect how suicidal expression propagates in digital communities. By employing a collaborative multi-agent approach, MACR aims to provide more nuanced and contextually rich risk assessments.
The technical architecture features two specialized AI agents working in tandem. The Reasoning Agent integrates cognitive appraisal theory to generate counterfactual user reactions to posts, scaling potential interactions and analyzing them through dedicated sub-agents for cognitive, emotional, and behavioral patterns. The Bias-aware Decision-Making Agent then uses a front-door adjustment strategy, leveraging these counterfactual reactions to mitigate harmful hidden biases. This collaboration allows the framework to both enrich contextual information and reduce bias, with extensive experiments on real-world conversational datasets demonstrating its effectiveness and robustness in identifying suicide risk, marking a significant step toward more sophisticated AI-powered mental health monitoring tools.
- Uses a two-agent system: Reasoning Agent generates counterfactual reactions; Bias-aware Agent mitigates hidden social biases.
- Addresses limitations of existing methods that miss complex user interactions and influences like conformity.
- Demonstrated effectiveness through experiments on real-world online conversation datasets.
Why It Matters
Offers a more nuanced AI tool for early suicide risk detection in online spaces, potentially improving digital mental health interventions.