Offering Collaborative-Like Recommendations When Data Is Sparse: The Case of Attraction-Weighted Information Filtering

Abstract

We propose a low-dimensional weighting scheme to map information filtering recommendations into more relevant, collaborative filtering-like recommendations. Similarly to content-based systems, the closest (most similar) items are recommended, but distances between items are weighted by attraction indexes representing existing customers’ preferences. Hence, the most preferred items are closer to all the other points in the space, and consequently more likely to be recommended. The approach is especially suitable when data is sparse, since attraction weights need only be computed across items, rather than for all user-item pairs. A first study conducted with consumers within an online bookseller context, indicates that our approach has merits: recommendations made by our attraction-weighted information filtering recommender system significantly outperform pure information filtering recommendations, and favorably compare to data-hungry collaborative filtering systems.