Deep Hierarchical Ensemble Model for Suicide Detection on Imbalanced Social Media Data

Area of Science:

• Computational social science
• Artificial intelligence
• Mental health research

Background:

• Suicide is a global crisis requiring timely intervention.
• Social media offers new avenues for suicide detection but faces challenges like data imbalance and implicit expressions.
• Existing methods struggle with the nuances of online user behavior related to suicidal ideation.

Purpose of the Study:

• To propose and validate a novel Deep Hierarchical Ensemble model for Suicide Detection (DHE-SD).
• To address data imbalance and expression implicitness in social media suicide detection.
• To enhance the applicability of suicide detection models across diverse user populations.

Main Methods:

• Development of a Deep Hierarchical Ensemble model for Suicide Detection (DHE-SD).
• Construction of a large-scale dataset from Sina Weibo (550K+ posts, 4521 users).
• Validation using a public Weibo dataset (7329 users) and a novel sentence-level mask mechanism.

Main Results:

• The DHE-SD model achieved state-of-the-art performance on both the constructed and public Weibo datasets.
• The sentence-level mask mechanism improved model robustness and applicability.
• The model effectively identified suicidal ideation even when baseline models' performance degraded.

Conclusions:

• The proposed DHE-SD model is highly effective for detecting suicidal ideation on social media platforms.
• The hierarchical ensemble strategy and mask mechanism offer significant improvements over existing methods.
• This research contributes a valuable tool for early intervention and mental health support via social media analysis.