Detecting self-injurious content and assessing sources of online support on YouTube and Twitter social networks

Student thesis: Doctoral Thesis


Self-harm is a significant public health issue typically common in young individuals. The behaviour is associated with significant mental health problems like anxiety, depression and eating disorder. The issue of self-harming behaviour has been viewed as the 'tip of the iceberg;' only a few self-harming individuals seek clinical support. Although people who self-harm tend to be isolated and remain less socialised in offline settings, it was found that most self-injurers are socially active in online spaces, especially on social media sites. Self-harming individuals voice their behaviours and seek help on social media by creating and sharing content with online users. Hence, the need to investigate such content is critical. Little is known about the nature of self-harm content on online social spaces and what distinguishes such content from non-harmful content. Also, our knowledge of digital sources of information for selfinjurers and their operational strategy on online social networks is insuffcient. While the literature reported that members of society constantly misjudge self-harming individuals, there is an inadequate understanding of the public perceptions and attitudes on digital networks concerning self-harm behaviour. The objective of this research was to gain a better knowledge of (1) how YouTube and Twitter users discuss self-harm behaviour, (2) the views and opinions of online members regarding self-harm, and (3) the strategy of support organisations disseminating self-harm related information on social networks. Additionally, this doctoral study aimed to propose and evaluate an automatic technique for detecting self-harm content in digital social spaces. The research investigation was performed to fill a knowledge gap, as past empirical studies analysing digital self-harm content on social media have primarily used qualitative techniques. While surveys and interviews are sources of data collection for many researchers in this field, this doctoral study sourced data from two popular social media sites (YouTube and Twitter). These platforms allow self-harming people to convey information that is hard to disclose in surveys or interviews. The study employed a mixed-methods approach and used state-of-the-art machine learning techniques to analyse the retrieved data. The analyses of self-harm content from the platforms revealed essential themes such as pro-self-harm, anti-self-harm and clean commentators. Additionally, this doctoral study uncovered the different opinions of users concerning self-harm across the examined platforms. A model was proposed using supervised machine learning techniques that automatically classify comments showing self-harm signs. The classification tasks were performed in binary and multi-class settings. The model based on the binary classification achieves higher performance (precision and recall) accuracy. Its performance outweighs that of the model built in a multi-class scenario. On the other hand, support organisations engaging with self-harming people on Twitter social networks exhibit different strategies while disseminating information to support positive well-being. Although the study recognises the limitations of utilising YouTube and Twitter data, the analysis illustrated how the platforms were used to communicate self-harm behaviours.
Date of Award20 Jan 2023
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorDiane Rasmussen Pennington (Supervisor) & Ian Ruthven (Supervisor)

Cite this