Fig 5. Time-variant user level suicide risk prediction using LSTM+CNN.

It comprises of an LSTM model to generate probabilities of a post ($p_0^i$), which is a sequence of word embeddings. Inlined CNN model that convolves over a sequence of post-level probabilities ([Pr(S)Pr(I)Pr(B)Pr(A)]) to predict user-level (u_i) suicide risk.