Suicide Ideation Prediction Through Deep Learning: An Integration of CNN and Bidirectional LSTM with Word Embeddings

Christianah T. Oyewale, Ayodeji O. J. Ibitoye, Joseph Damilola Akinyemi, Olufade F. W. Onifade

Abstract: Suicide ideation prediction received attention in public health because it affects not only the victim but the general public. Identifying suicide ideation can be traditionally tricky and time-consuming, and the victim might have harmed himself or herself before detection. Easy access to social media provided textual data for Deep Learning (DL) models like Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for the identification of suicide ideation. These models are more promising at identifying suicide ideation than traditional Machine Learning models because they can be applied to every form of datasets. The main challenge is combining the DL models and word embeddings. This research employed the DL model by combining CNN with bidirectional LSTM and two-word embedding techniques, FastText and GloVe (global vectors for word representation). The research questions are as follows: Will there be a difference in F1 score between two-word embeddings when combined with CNN-BiLSTM? And will the overfitting issue in the existing study be addressed? Using an existing study as a baseline, experimentation was done using a secondary Reddit dataset of 232,074 posts, achieving a 94% F1-score for FastText and GloVe. It was observed that GloVe could do better without approximating the F1-score and FastText when the running time is considered.