DynaProtect : a dynamic factor influence learning framework for protective factor-aware suicide risk prediction

Abstract

Despite significant advances in approaches to suicide detection on social media, predicting users' suicide risk in a subsequent state remains challenging. Even though existing works have identified various risk factors to improve detection performance, they often overlook the critical role of protective factors in suicide prevention. To address this limitation, we propose an approach that jointly learns both risk and protective factors to predict users' subsequent suicide risk. Recognizing that the effectiveness of these factors varies across different user patterns, we introduce a dynamic factor influence learning mechanism that captures user-dependent interactions with risk and protective factors. Our experiments demonstrate that the integrated approach significantly enhances suicide risk prediction performance compared to existing methods.