Propagating Maximum Capacities for Recommendation

Neighborhood-based approaches often fail in sparse scenarios; a direct implication for recommender systems exploiting co-occurring items is often an inappropriately poor performance. As a remedy, we propose to propagate information (e.g., similarities) across the item graph to leverage sparse data. Instead of processing only directly connected items (e.g. co-occurrences), the similarity of two items is defined as the maximum capacity path interconnecting them. Our approach resembles a generalization of neighborhood-based methods that are obtained as special cases when restricting path lengths to one. We present two efficient online computation schemes and report on empirical results.