Suicidality Detection on Social Media Using Metadata and Text Feature Extraction and Machine Learning

Abstract

In this study, we implemented machine learning models that can detect suicidality posts on Twitter. We randomly selected and annotated 20,000 tweets and explored metadata and text features to build effective models. Metadata features were studied in great details to understand their possibility and importance in suicidality detection models. Results showed that posting type (i.e., reply or not) and time-related features such as the month, day of the week, and the time (AM vs. PM) were the most important metadata features in suicidality detection models. Specifically, the probability of a social media post being suicidal is higher if the post is a reply to other users rather than an original tweet. Moreover, tweets created in the afternoon, on Fridays and weekends, and in fall have higher probabilities of being detected as suicidality tweets compared with those created in other times. By integrating metadata and text features, we obtained a model of good performance (i.e., F1 score of 0.846) that can assist humans in the real-world setting to detect suicidality social media posts.

Keywords:

• Classification
• feature extraction
• machine learning
• social media
• suicidality detection

Additional information

Funding

This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ICT Creative Consilience program [IITP-2020-0-01821) supervised by the IITP (Institute for Information & communications Technology Planning & Evaluation); Ministry of Education of the Republic of Korea and the National Research Foundation of Korea [NRF-2019S1A5A2A03045289]; and SKKU-SMC Future Convergence Research Program Grant.

Notes on contributors

Woojin Jung

Woojin Jung, Donghun Kim, Seojin Nam, and Yongjun Zhu, Department of Library and Information Science, Sungkyunkwan University, Seoul, Republic of Korea

Donghun Kim

Woojin Jung, Donghun Kim, Seojin Nam, and Yongjun Zhu, Department of Library and Information Science, Sungkyunkwan University, Seoul, Republic of Korea

Seojin Nam

Woojin Jung, Donghun Kim, Seojin Nam, and Yongjun Zhu, Department of Library and Information Science, Sungkyunkwan University, Seoul, Republic of Korea

Yongjun Zhu

Woojin Jung, Donghun Kim, Seojin Nam, and Yongjun Zhu, Department of Library and Information Science, Sungkyunkwan University, Seoul, Republic of Korea