Generating Social Network Features for Link-based Classification Jun Karamon 1 , Yutaka Matsuo 2 , Hikaru Yamamoto 3 , and Mitsuru Ishizuka 1 1 The University of Tokyo, 7-3-1 Hongo Bunkyo-ku Tokyo, Japan, karamon@mi.ci.i.u-tokyo.ac.jp,ishizuka@i.u-tokyo.ac.jp 2 National Institute of Advanced Industrial Science and Technology, 1-18-13 Soto-kanda, Chiyoda-ku, Tokyo, Japan, y.matsuo@aist.go.jp 3 Seikei University, 3-3-1 Kichijoji Kitamachi, Musashino-shi, Tokyo, Japan yamamoto@econ.seikei.ac.jp Abstract. There have been numerous attempts at the aggregation of attributes for relational data mining. Recently, an increasing number of studies have been undertaken to process social network data, partly be- cause of the fact that so much social network data has become available. Among the various tasks in link mining, a popular task is link-based classification, by which samples are classified using the relations or links that are present among them. On the other hand, we sometimes employ traditional analytical methods in the field of social network analysis us- ing e.g., centrality measures, structural holes, and network clustering. Through this study, we seek to bridge the gap between the aggregated features from the network data and traditional indices used in social network analysis. The notable feature of our algorithm is the ability to invent several indices that are well studied in sociology. We first define general operators that are applicable to an adjacent network. Then the combinations of the operators generate new features, some of which cor- respond to traditional indices, and others which are considered to be new. We apply our method for classification to two different datasets, thereby demonstrating the effectiveness of our approach. 1 Introduction Recently, increasingly numerous studies have been undertaken to process net- work data (e.g., social network data and web hyperlinks), partly because of the fact that such great amounts of network data have become available. Link min- ing [6] is a new research area created by the intersection of work in link analysis, hypertext and web mining, relational learning, and inductive logic programming and graph mining. A popular task in link mining is link-based classification, classifying samples using the relations or links that are present among them. To date, numerous approaches have been proposed for link-based classification [13], which are often applied to social network data. A social network is a social structure comprising nodes (called actors) and relations (called ties). Prominent examples of recently studied social networks