Recommending Journal Articles with PageRank Ratings

The TechLens+ strategy for addressing the recommender cold-start problem in a scholarly digital library is to seed the preference matrix with article references. However, this method generates boolean ratings rather than ratings on a numerical scale, as is more typical with recommender systems for commodity products. One strategy for generating a better preference matrix for collaborative filtering recommendations is to compute the PageRank values for the articles in the citation graph of the article collection and to substitute the boolean ratings with PageRank “ratings”. There is a significant amount of prior research which suggests that this strategy should generate better Top-N recommendations. However, the experimental results described in this paper show that PageRank ratings are inferior to both boolean ratings and random (but consistent) ratings.