In this paper, we consider the problem of user modeling in online social networks, and propose a social interaction activity based user vectorization framework, called the time-varying user vectorization (Tuv), to infer and make use of important user features. Tuv is designed based on a novel combination of word2vec, negative sampling and a smoothing technique for model training. It jointly handles multi-format user data and computes user representing vectors, by taking into consideration user feature variation, self-similarity and pairwise interactions among users. The framework enables us to extract hidden user properties and to produce user vectors. We conduct extensive experiments based on a real-world dataset, which show that Tuv significantly outperforms several state-of-the-art user vectorization methods.