Abstract
In a growing number of relational domains, the data record temporal sequences of interactions among entities. For example, in citation domains authors publish scientific papers together each year and in telephone fraud detection domains people make calls to each other each day. The temporal dynamics of these interactions contain information that can improve predictive models (e.g., people publishing together frequently are likely to be publishing on the same topic) but to date there has been little effort to incorporate timevarying dependencies into relational models. Past work in relational learning has focused primarily on static "snapshots" of relational data. In this paper, we present an initial approach to modeling dynamic relational data graphs in predictive models of attributes. More specifically, we use a two-step process that first summarizes the dynamic graph with a weighted static graph and then incorporates the link weights in a relational Bayes classifier. We evaluate our approach on the Cora dataset (where co-author and citation links vary over time) showing that our approach results in significant performance gains over a baseline snapshot approach that ignores the temporal component of the data.
