Guy Shani, David Maxwell Chickering, and Christopher Meek
In this paper we show how to use data from the internet to construct recommender systems. We provide experimental results, which include a user study, showing that our methods produce good recommendations in realistic applications. We describe how standard evaluation metrics are biased toward systems that simply recommend the most popular items, and we propose modifications to these metrics to fix the problem.
In ACM International Conference on Recommender Systems