Social monitoring for risk prediction in public forums

• Autor(a) principal: Barros, Lucas Filipe Roberto de
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/34066
• topic: Social media; Hyperparameters; Self-harm; Machine learning

In the last couple of years, Social media has been changing the methods people use to communicate and express themselves. One big difference between this new communication mechanism and the traditional methods is that this data is documented. Hence, it is important to understand how all of this data can be used to improve our lives. This thesis main subject is to use this data to perceive if a user shows any signs of mental health problems. We also participated, with the group named Bioinformatics of the Institute of Electronics and Computer Engineering of University of Aveiro (BioInfo@UAVR), in the 2nd shared task of CLEF eRisk 2021. eRisk is an “Early Risk Prediction on the Internet” online challenge whose tasks consist in analysing social media data and foster research on early detection of mental disorders. This year eRisk had 3 tasks, each focusing on a different disorder. This paper focuses on addressing the 2nd task, whose main objective is the early detection of users at risk of self-harming, based on their history. This issue was addressed by developing supervised machine learning models that can classify such users. In this approach, we used Tokenization algorithms based on regular expressions and Yake (which is a keyword extractor tool), linguistic features such as emojis, and other models such as BERT embeddings and VADER sentiment score and machine learning classifiers such as Support Vector Machines or Boosting classifiers. After testing all combinations of methods, with different combinations of hyperparameters, we conclude that it is possible to make decisions regarding a user’s mental health state based on these methods with latency − weighted score of 0.46 for the eRisk 2021 testset.