Temporal convolutional networks with language models for decoding music preferences in mental health profiling

Music preferences serve as crucial behavioural clues for decoding mental health states.Music provides a continuous and emotionally rich behavioural signal that is less influenced by social desirability biases compared to self-reported data, making it a robust indicator for mental health assessment.However, traditional analysis methods struggle to simultaneously account for the temporal dynamics of music listening and its rich semantic information, resulting in limited decoding efficacy.Previous studies attempted hybrid models but often faced overfitting or computational inefficiency, which motivated our design of a more integrated framework.To address this, we propose an innovative framework that integrates temporal convolutional networks with pre-trained language models to capture both the sequential patterns of music consumption and the emotional semantics of lyrics content.Our validation on a public dataset containing over 100,000 records demonstrates that this model achieves approximately 8.5% higher accuracy than single-modal benchmark methods in mental health state assessment tasks.It also effectively identifies specific musical features associated with depressive and anxious tendencies.This work provides a novel technical pathway for achieving non-invasive, dynamic mental health screening.