Personal Credit Profiling via Latent User Behavior Dimensions on Social Media

Consumer credit scoring and credit risk management have been the core research problem in financial industry for decades. In this paper, we target at inferring this particular user attribute called credit, i.e., whether a user is of the good credit class or not, from online social data. However, existing credit scoring methods, mainly relying on financial data, face severe challenges when tackling the heterogeneous social data. Moreover, social data only contains extremely weak signals about users’ credit label. To that end, we put forward a Latent User Behavior Dimension based Credit Model (LUBD-CM) to capture these small signals for personal credit profiling. LUBD-CM learns users’ hidden behavior habits and topic distributions simultaneously, and represents each user at a much finer granularity. Specifically, we take a real-world Sina Weibo dataset as the testbed for personal credit profiling evaluation. Experiments conducted on the dataset demonstrate the effectiveness of our approach: (1) User credit label can be predicted using LUBD-CM with a considerable performance improvement over state-of-the-art baselines; (2) The latent behavior dimensions have very good interpretability in personal credit profiling.