Time-Evolving Relational Classification and Ensemble Methods Ryan Rossi and Jennifer Neville Purdue University, West Lafayette, IN 47906, USA {rrossi,neville}@purdue.edu Abstract. Relational networks often evolve over time by the addition, deletion, and changing of links, nodes, and attributes. However, accu- rately incorporating the full range of temporal dependencies into relational learning algorithms remains a challenge. We propose a novel framework for discovering temporal-relational representations for classi- fication. The framework considers transformations over all the evolving relational components (attributes, edges, and nodes) in order to accu- rately incorporate temporal dependencies into relational models. Addi- tionally, we propose temporal ensemble methods and demonstrate their effectiveness against traditional and relational ensembles on two real- world datasets. In all cases, the proposed temporal-relational models outperform competing models that ignore temporal information. 1 Introduction Temporal-relational information is present in many domains such as the Internet, citation and collaboration networks, communication and email networks, social networks, biological networks, among many others. These domains all have at- tributes, links, and/or nodes changing over time which are important to model. We conjecture that discovering an accurate temporal-relational representation will disambiguate the true nature and strength of links, attributes, and nodes. However, the majority of research in relational learning has focused on mod- eling static snapshots [2, 6] and has largely ignored the utility of learning and incorporating temporal dynamics into relational representations. Temporal relational data has three main components (attributes, nodes, links) that vary in time. First, the attribute values (on nodes or links) may change over time (e.g., research area of an author). Next, links might be created and deleted throughout time (e.g., host connections are opened and closed). Finally, nodes might appear and disappear over time (e.g., through activity in an online social network). Within the context of evolving relational data, there are two types of predic- tion tasks. In a temporal prediction task, the attribute to predict is changing over time (e.g., student GPA), whereas in a static prediction task, the predic- tive attribute is constant (e.g., paper topic). For these prediction tasks, the space of temporal-relational representations is defined by the set of relational P.-N. Tan et al. (Eds.): PAKDD 2012, Part I, LNAI 7301, pp. 1–13, 2012. c  Springer-Verlag Berlin Heidelberg 2012