Detecting Malicious Users in Online Dating Application

The flood of malicious users whose baleful behaviors seriously affect the social experience of others leads to the loss of common users. Currently, the detection and control of malicious users in mobile social applications mainly depend on users' reports, which is not timely. In this paper, we present the approaches to identify the malicious users according to information integrity, behavior characteristics and conversation content in early time. Considering statistical methods, we propose the Time Series (TIME) model based on the analysis of users' time behavior characteristics. And thinking from the theme of the users' session, the improved Latent Dirichlet Allocation (LDA) malicious topic model is designed which has improved the sensitivity of the model to detect malicious topics. To ensure the feasibility and accuracy of the models, we compare our models with the traditional machine learning methods of Decision Tree (DEC) and Support Vector Machine (SVM) in online dating application. The final evaluation results have demonstrated that the precision of our models has exceeded the traditional model over 8% in the identification of malicious users.