Understanding evolution of research themes: a probabilistic generative model for citations
Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining
| Hi-index | 0.00 |
Many algorithms have been developed to identify important nodes in a complex network, including various centrality metrics and Page Rank, but most fail to consider the dynamic nature of the network. They therefore suffer from recency bias and fail to recognize important new nodes that have not had as much time to accumulate links as their older counterparts. This paper describes the Effective Contagion Matrix (ECM), a solution to address recency bias in the analysis of dynamic complex networks. The idea of ECM is to explicitly consider the temporal order of links and chains of links connecting to a node with some temporal decay factors. We tested ECM with three large real world citation networks on the task of predicting papers' future importance. We compared ECM's performance with two static metrics, degree-centrality and Page Rank, and two time-aware metrics, age-based Page Rank and Cite Rank. We show that ECM is more appropriate for predicting future citations and Page Rank scores with regard to new citations. We also describe a procedure to estimate ECM's parameters from the data. Combining all five scores into a v-SVR regression model of future citations improves the predictive performance further.