Deep Learning Approaches for Analyzing the Impact of Social Media Usage on Young Students: A Systematic Literature Review

Abstract

Social media (SM) is ubiquitous to today's school-age and college-age students, and the continuous creation of text, image and behavioral data is used more and more for deep learning to understand its relationship to mental health, behavior and academic life. This systematic review aims to summarise the evidence on the use of deep learning approaches to analyse the effects of social media data on young students. According to the PRISMA 2020 protocol, a total of 214 reports were identified in the academic databases or preprint repositories, of which 21 were duplicates, 193 were screened at the title and abstract level, 72 reports were sought for retrieval, 70 reports were assessed at the full text stage and 30 studies were included in the qualitative synthesis. The domains covered were primarily involving the detection of addiction and problematic-use, detection of depression and suicidal-ideation, detection of cyberbullying, detection of misinformation, and detection of stress as well as sentiment/emotion recognition, with the former being implemented primarily using the CNN-RNN hybrid, and the latter using the transformer/BERT-based architecture. Some of the best reported classification rates were, around 71%, to very high rates of above 99%, such as an CNN–LSTM ensemble with attention that reported a high 90.3% for detecting suicidal ideation [17] and a high 95.6% accuracy using a convolutional recurrent network for student addiction detection.