Evaluating Publication Similarity Measures

Publication searching based on keywords provided by users is traditional in digital libraries. While useful in many circumstances, the success of locating related publications via keyword-based searching paradigm is influenced by how users choose their keywords. Example-based searching, where user provides an example publication to locate similar publications, is also becoming commonplace in digital libraries. Existing publication similarity measures, needed for example-based searching, fall into two classes, namely, text-based similarity measures from Information Retrieval, and citation-based similarity measures based on bibliographic coupling and/or co-citation. In this paper, we list a number of publication similarity measures, and extend and evaluate them in terms of their accuracy, separability, and independence. For evaluation, we use the ACM SIGMOD Anthology, a digital library of about 15,000 publications.