Leveraging deep learning model to visually detect changes in depression tendency of social media users

Purpose Leveraging deep learning models to detect the depressive tendencies of social media users and visualizing the trajectories of their emotional changes. Addressing the issues of subjectivity bias and high implementation cost existing in traditional depression detection approaches. Design/methodology/approach A new multi-instance Long short-term memory (LSTM) learning model is put forward. It employs an autoencoder to extract time-series features from the text, constructs a binary classifier to sift depression-related tweets, determines the depression tendencies of users and generates visual graphs based on the distribution of the time series data to exhibit the dynamic variations of users’ emotions. Findings Our model is capable of outputting the probabilities of depressive tendencies for each of the user’s tweets, and its performance is stronger than that of the fundamental methods. Practical implications Our study achieves the dynamic and visual monitoring of depressive tendencies, breaking through the constraints of traditional single-time-point detection. Utilizing social media text data, a low-cost and scalable large-scale depression screening solution is proposed. Originality/value Our study achieves the dynamic and visual monitoring of depressive tendencies, breaking through the constraints of traditional single-time-point detection. Utilizing social media text data, a low-cost and scalable large-scale depression screening solution is proposed.