Incorporating historical information by disentangling hidden representations for mental health surveillance on social media

The growing need to identify mental health conditions has paved the way for automated computational methods for mental health surveillance on social media. However, inferring the accurate state of a user’s mind requires understanding the history of the user’s mental health condition, which is critical for identifying the mental health landscape of at-risk users. Recent methods have offered performance improvement to address this challenging area; however, they assume that the intervals between all historical social media posts are equally important, making them unable to capture the dynamic temporal patterns of historical posts. In this work, we address this gap and incorporate a time-aware framework that jointly learns the context of the user’s historical posts and temporal posting irregularities by disentangling representations of different time intervals in the users’ historical posts and adaptively selecting the most important interval for each sample at each time step. First, the hidden state of the RNN is disentangled into multiple independently updated small hidden states to model users’ historical posting information. Then, at each time step, the temporal context information is used to modulate the features of different posts, selecting the most important interval within the historical posts. Experimental results on two mental health (i.e., depression and self-harm) Reddit datasets show that our method outperforms state-of-the-art methods for mental health surveillance on social media.