A statistical infinite feature cascade-based approach to anomaly detection for dynamic social networks

The development of methods for anomaly detection in dynamic ubiquitous online social networks is critical to coincide with the growth in social network usage. This paper presents a novel statistical approach to anomaly detection in dynamic social networks. The approach relies upon the fact that the network dynamics can be driven by microscopic features of each node that dynamically cascade to neighboring nodes over time. The proposed approach consists of two main components: (1) normal modeling component and (2) anomaly detection component. The former component is involved in three main processes, governing the network dynamics. The first process is the features' birth, death, and lifetime, which is assumed to follow a realistic statistical distribution in this paper for the very first time. The second process is the evolution of nodes' features that is modeled by an Infinite Factorial Hidden Markov Model (IFHMM), considering feature cascade. The feature cascade is a phenomenon that explicitly describes how the past features of each node affect the features of its neighboring nodes in future. The third process modeled in this paper is the relationship between nodes' features and link generation in dynamic social networks. The latter component of the proposed approach provides a new method to quantize deviation of network dynamics from the normal behavior. Some Markov Chain Monte Carlo (MCMC) sampling strategies have been used to simulate parameters of the proposed model, given social network data. The proposed anomaly detection approach is validated by experiments on synthetic and real social network datasets. Experimental results show that this approach outperforms other related approaches in terms of some statistical performance measures, especially applied to binary normal-abnormal classification test.