Reviewer Profiling Using Sparse Matrix Regression

Thousands of scientific conferences happen every year, and each involves a laborious scientific peer review process conducted by one or more busy scientists serving as Technical/Scientific Program Committee (TPC) chair(s). The chair(s) must match submitted papers to their reviewer pool in such a way that i) each paper is reviewed by experts in its subject matter, and ii) no reviewer is overloaded with reviews or under-utilized. Towards this end, seasoned TPC chairs know the value of reviewer and paper profiling: summarizing the expertise/interests of each reviewer and the subject matter of each paper using judiciously chosen domain-specific keywords. An automated profiling algorithm is proposed for this purpose, which starts from generic/noisy reviewer profiles extracted using Google Scholar and derives custom conference-centric reviewer and paper profiles. Each reviewer is expert on few sub-topics, whereas the pool of reviewers and the conference may collectively need many more keywords for appropriate specificity. Exploiting this sparsity, we propose a sparse matrix factorization approach in lieu of classical SVD-based LSI or NMF-type approaches. We illustrate the merits of our approach using real conference data, and expert scoring of the assignments by a seasoned TPC chair in the area.