TRANSACTION ON KNOWLEDGE AND DATA ENGINEERING, 2010 1 Identifying Evolving Groups in Dynamic Multi-Mode Networks Lei Tang, Member, IEEE, Huan Liu, Senior Member, IEEE, and Jianping Zhang Abstract—A multi-mode network consists of heterogeneous types of actors with various interactions occurring between them. Identifying communities in a multi-mode network can help understand the structural properties of the network, address the data shortage and unbalanced problems, and assist tasks like targeted marketing and finding influential actors within or between groups. In general, a network and its group structure often evolve unevenly. In a dynamic multi- mode network, both group membership and interactions can evolve, posing a challenging problem of identifying these evolving communities. In this work, we try to address this problem by employing the temporal information to analyze a multi-mode network. A temporally-regularized framework and its convergence property are carefully studied. We show that the algorithm can be interpreted as an iterative la- tent semantic analysis process, which allows for extensions to handle networks with actor attributes and within-mode interactions. Experiments on both synthetic data and real- world networks demonstrate the efficacy of our approach and suggest its generality in capturing evolving groups in networks with heterogeneous entities and complex relationships. Index Terms—Data Mining, Community Detection, Com- munity Evolution, Multi-Mode Networks, Dynamic Networks I. I NTRODUCTION O WING to the widely-available network data produced from social networks, technology networks, informa- tion networks and genetic regulatory networks [2], network analysis [3] and modeling [4] is attracting increasing atten- tion from many fields. Examples include epidemiology [5], intelligence analysis [6], targeted marketing, recommen- dation systems [7], relational learning [8] and behavior prediction [9]. A large body of existing work deals with networks of one mode. That is, only one type of actors (nodes) are present in a network, and the connections (interactions) between actors are of the same type. This is common for friendship networks, and mobile networks. Recently, burgeoning applications such as web mining, collaborative filtering, and online targeted marketing involve more than one type of entities. Between them are different types of interaction. This kind of network is called multi-mode network [3] (a.k.a. heterogeneous network). A preliminary version of this work is published in [1]. L.Tang and H. Liu are with Computer Science & Engineering, Arizona State University, Tempe, AZ, 85287, USA. J. Zhang is with the MITRE Corporation, McLean, VA, 22102, USA. E-mail: L.Tang@asu.edu, huanliu@asu.edu, jzhang@mitre.org          Fig. 1. An example of 3-Mode Network in YouTube A. Multi-Mode Networks Take the network in YouTube as an example. A 3-mode network (shown in Figure 1) can be constructed: users, videos, and tags. Note that in this network, both videos and tags are considered “actors” as well, although the user might be the major mode under consideration. Different interactions exist between the three types of entities: users can upload videos; users can add tags to a video. Videos and tags are naturally correlated to each other. Meanwhile, a friendship network exists between users, and a video clip can be uploaded to respond to another video. Tags can also connect to each other based on their semantic meanings. In other words, multiple types of entities exist in the same network, and entities relate to others (either the same type or different types) through different links. Another example of multi-mode network is the field of academia as shown in Figure 2. Assorted entities (researchers, conference/journals, papers, words) are in- tertwined with each other. Scientific literature connects papers by citations; papers are published at different places (conferences, journals, workshops, thesis, etc.); and re- searchers are connected to papers through authorship. Some might also relate to each other by serving simultaneously as journal editors or on conference program committees. Moreover, each paper can focus on different topics repre- sented by words. Words are associated to each other based on semantics. At the same time, papers are connected to different conferences and journals. Within a multi-mode network, different types of entities tend to form groups or communities 1 . In the YouTube example, users sharing similar interests are more likely to form a group; videos are clustered naturally if they relate to similar contents; and tags are clustered if they are associated with similar users and videos. Generally, a user group may 1 Group and community are interchangeable in this work.