Pairwise structural role mining for user categorization in information cascades

Sarvenaz Choobdar, Pedro Ribeiro, Fernando Silva
2015 Proceedings of the 2015 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining 2015 - ASONAM '15  
The tendency of users to connect with peers of similar interests and social demography (homophily) is one of the sources of information for user behavior modeling and classification. However this is yet an open question for structural roles where nodes at similar structural position in the network play the same roles: are structurally equivalent nodes more prone to have connections between themselves? In this paper, we tackle this open question by studying the patterns of homophily for
more » ... l roles. We propose a new method named SR-Diffuse to simultaneously identify structural roles in a network and to model the role membership matrix of users. In this method, we integrate pairwise role dependency alongside with structural features of users for role mining. We show that pairwise role dependency is necessary to distinguish some structural roles but it is a misleading factor for some others. We design an optimization model to capture structural roles with the guidance of pairwise dependency, and devise an iterative algorithm to learn structural roles simultaneously from structural properties and social dependency of users. We examine the efficacy of our new method in a users classification problem for information cascades. We compare the predictability of discovered roles by our method against some baseline methods for predicting social classes of users in different information cascades in two social networks, Flickr and Digg. The experimental results suggest that our method can improve the quality of roles membership of users and can better represent the profile of users in the network, hence it is a better predictor for social classes of users in an information cascade.
doi:10.1145/2808797.2808909 dblp:conf/asunam/ChoobdarRS15 fatcat:q2rptkeahvdmtfrfgxvpt6xq6m