Efficient Privacy-Preserving Collaborative Filtering Based on the Weighted Slope One Predictor

Rating-based collaborative filtering (CF) predicts the rating that a user will give to an item, derived from the ratings of other items given by other users. Such CF schemes utilise either user neighbourhoods (i.e. user-based CF) or item neighbourhoods (i.e. item-based CF). Lemire and MacLachlan [19] proposed three related schemes for an item-based CF with predictors of the form f (x) = x+b, hence the name “slope one”. Slope One predictors have been shown to be accurate on large datasets. They also have several other desirable properties such as being updatable on the fly, efficient to compute, and work even with sparse input. In this paper, we present a privacy-preserving item-based CF scheme through the use of an additively homomorphic public-key cryptosystem on the weighted Slope One predictor; and show its applicability on both horizontal and vertical partitions, and include a discussion on arbitrary partitions as well. We present an evaluation of our proposed scheme in terms of communication and computation complexity, performance of cryptographic primitives and performance of a single-partition, single machine implementation in 64-bit Java.