III: Small: Dynamic Social Network Mining: Feature Extraction, Modeling and Anomaly Detection

Suny At Buffalo, Amherst NY

Investigators

Abstract

This project develops a general framework for anomaly detection in dynamic social networks that evolve in both links and nodes. The framework includes first capturing the information of timestamps of dynamic networks by transferring them into carefully selected graph kernel feature spaces. A dynamic modeling method is then designed to learn the dynamism on the target dynamic social network. Anomaly detection methods are finally developed to mine abnormal nodes in the dynamic network. The main innovation of the approaches is to represent dynamic networks by bags of attributes, including graph kernel features to epitome details in each timestamp, learned latent variables for dynamism of networks and user specified features to turn the direction of attributes representation towards aimed tasks. Based on this representation, the project designs innovative methods based on latent support vector machine and transfer learning to detect abnormal nodes. This research provides a clear understanding of evolution patterns, including both normality and abnormality in dynamic social networks. The approaches developed in this project help identify various abnormalities in our life, including detecting spammers in websites, monitoring potential dangerous activities in crime networks, identifying malicious source of infection in disease networks, and many others. The successful modeling of such network dynamics can provide scientific basis for appearance and disappearance of human relationships, improve the prospects for uncovering potentially undiscovered evolution patterns in social networking process and help develop qualitative and quantitative algorithms for more applications.