Real-time recommendation with locality sensitive hashing

Neighborhood-based collaborative filtering (CF) methods are widely used in recommender systems because they are easy-to-implement and highly effective. One of the significant challenges of these methods is the ability to scale with the increasing amount of data since finding nearest neighbors requires a search over all of the data. Approximate nearest neighbor (ANN) methods eliminate this exhaustive search by only looking at the data points that are likely to be similar. Locality sensitive hashing (LSH) is a well-known technique for ANN search in high dimensional spaces. It is also effective in solving the scalability problem of neighborhood-based CF. In this study, we provide novel improvements to the current LSH based recommender algorithms and make a systematic evaluation of LSH in neighborhood-based CF. Besides, we make extensive experiments on real-life datasets to investigate various parameters of LSH and their effects on multiple metrics used to evaluate recommender systems. Our proposed algorithms have better running time performance than the standard LSH-based applications while preserving the prediction accuracy in reasonable limits. Also, the proposed algorithms have a large positive impact on aggregate diversity which has recently become an important evaluation measure for recommender algorithms.