PageRank Effect on Collaborative Filtering July 30, 2008Posted by Andre Vellino in Collaborative filtering, Digital library, Recommender.
I have done some experiments on the impact of PageRank on a collaborative filtering recommender for journal articles. The results are counterintuitive – to me anyway – but I think they might have a plausible explanation (I’m working on one anyway.)
I followed in the footsteps of TechLens+ and used article references as a proxy for “ratings” – in other words, assume that one article citing another means a (boolean) “positive vote” for the cited article. It’s a poor approximation, but it addresses the cold-start problem for a digital library recommender.
The idea behind using PageRank was to refine these boolean ratings and rank them on a scale. Using numeric PageRank values on the ratings (rather than a boolean value) has a surprising effect: Top-N prediction quality goes down! Furthermore, random values for PageRank are about the same as boolean (constant) values for PageRank.
I trust the Daniel Lemire is right about the value of negative results.