Detecting Traces of Self-harm in Social Media: A Simple and Interpretable Approach

Juan Aguilera, Delia Irazú Hernández Farías, Manuel Montes-y-Gómez, Luis Carlos González-Gurrola

Abstract: Social networks have become the main means of communication and interaction between people. In them, users share information and opinions, but also their experiences, worries, and personal concerns. Because of this, there is a growing interest in analyzing this kind of content to identify people who commit self-harm, which is often one of the first signs of suicide risk. Recently, methods based on Deep Learning have shown good results in this task, however, they are opaque and do not facilitate the interpretation of decisions, something fundamental in health-related tasks. In this paper, we face the detection of self-harm in social media by applying a simple and interpretable one-class-classification approach, which, supported on the concept of the attraction force [1], produces its decisions considering both the relevance and distance between users. The results obtained in a benchmark dataset are encouraging, as they indicate a competitive performance with respect to state-of-the-art methods. Furthermore, taking advantage of the approach’s properties, we outline what could be a support tool for healthcare professionals for analyzing and monitoring self-harm behaviors in social networks.