Prediction model of social network suicide ideation by small sample

Abstract

Social media is one of the platforms for many people to express their sentiments. However, more and more post on social media contain suicide information. At present, the rapid generation of online suicide post cannot be balanced with the speed of manual annotation, which makes it very difficult to establish a fast and real-time predictive model. In order to identify suicidal actors on social media, this study proposes a mechanism to build a small sample of suicidal ideation prediction model. Through this mechanism, the model selects appropriate sample information and obtains accurate annotation data at the lowest cost. Natural language processing technology is quite complex. In order to ensure the quality of the text representation of the input value, this study adopts the BERT with multi-head attention mechanism as the text representation extraction model. Considering that social media content is usually used to express personal sentiments, the text representation must be weighted by a layer of text sentiment representation. We use the Random Forest of the ensemble learning concept as the classifier. In order to ensure that the model uses the most informative data, this research also uses active learning. Our proposed architecture has achieved exciting results on small data sets. The random forest model is paired with an active learning mechanism to train weighted and unweighted text representation, and then make predictions on the test set. The results show that the combination of weighted text representation and active learning can improve the accuracy of the baseline classifier by 10%. The framework proposed in this study can increase the accuracy of the baseline model and reduce the labor and time costs of a large amount of data labeling. It is expected to help suicide prevention and control related units to explore the difficulty of the Internet.