Co-Bidding Graphs for Constrained Paper Clustering

The information for many important problems can be found in various formats and modalities. Besides standard tabular form, these include also text and graphs. To solve such problems fusion of different data sources is required. We demonstrate a methodology which is capable to enrich textual information with graph based data and utilize both in an innovative machine learning application of clustering. The proposed solution is helpful in organization of academic conferences and automates one of its time consuming tasks. Conference organizers can currently use a small number of software tools that allow managing of the paper review process with no/little support for automated conference scheduling. We present a two-tier constrained clustering method for automatic conference scheduling that can automatically assign paper presentations into predefined schedule slots instead of requiring the program chairs to assign them manually. The method uses clustering algorithms to group papers into clusters based on similarities between papers. We use two types of similarities: text similarities (paper similarity with respect to their abstract and title), together with graph similarity based on reviewers’ co-bidding information collected during the conference reviewing phase. In this way reviewers’ preferences serve as a proxy for preferences of conference attendees. As a result of the proposed two-tier clustering process similar papers are assigned to predefined conference schedule slots. We show that using graph based information in addition to text based similarity increases clustering performance. The source code of the solution is freely available. 1998 ACM Subject Classification I.2.8 Problem Solving, Control Methods, and Search