Subjective Collaborative Filtering

TL;DR

This paper introduces an efficient item-based collaborative filtering method that uses approximate Jaccard similarity and user preferences to improve recommendation accuracy and computational efficiency.

Contribution

It proposes a novel approximation of item similarity using constant-time operations and incorporates user preferences for enhanced recommendation quality.

Findings

Efficient similarity approximation using bitwise operations

Incorporation of user preferences improves recommendation relevance

Method achieves fast computation suitable for large datasets

Abstract

We present an item-based approach for collaborative filtering. We determine a list of recommended items for a user by considering their previous purchases. Additionally other features of the users could be considered such as page views, search queries, etc... In particular we address the problem of efficiently comparing items. Our algorithm can efficiently approximate an estimate of the similarity between two items. As measure of similarity we use an approximation of the Jaccard similarity that can be computed by constant time operations and one bitwise OR. Moreover we improve the accuracy of the similarity by introducing the concept of user preference for a given product, which both takes into account multiple purchases and purchases of related items. The product of the user preference and the Jaccard measure (or its approximation) is used as a score for deciding whether a given product has to be recommended.