Profiling human behavior through multidimensional latent factor modeling

The Human Behavioral Analysis is a growing research area due to its big impact on several scientific and industrial applications. One of the most popular family of techniques addressing this problem is the Latent Factor Modeling which aims at identifying interesting features that determine human behavior. In most cases, latent factors are used to relate atomic features to each other: for example, semantically similar words in documents of a textual corpus (text analysis), products to buy and customers (recommendation), users (social influence) or news in a social network (information diffusion). In this paper, we propose a new latent-factor-based approach whose goal is to profile users according to their behavior. The novelty of our proposal consists in considering the actions as set of features instead of single atomic elements. A single action is characterized by several components that can be exploited in order to define fine-grain user profiles. These components can be, for instance, “what is being done”, “where”, “when” or “how”.We evaluated our approach in two application scenarios. A first test is performed on real data and it is aimed at semantically validate the model identifying behavioral clusters of users; a second test is a predictive experiment on synthetic data generated to assess model’s anomaly detection capability.