Hybrid recommendation methods in complex networks

TL;DR

This paper introduces two novel hybrid recommendation algorithms that normalize similarity measures and combine user and item scores, demonstrating up to 20% performance improvements across various datasets.

Contribution

The authors propose new recommendation methods based on normalized similarity measures and score combination, validated on multiple datasets with comparative analysis.

Findings

Up to 20% performance improvement over existing methods

Recommendation accuracy varies significantly across different networks

One algorithm outperforms others in noisy, link-rich datasets

Abstract

We propose here two new recommendation methods, based on the appropriate normalization of already existing similarity measures, and on the convex combination of the recommendation scores derived from similarity between users and between objects. We validate the proposed measures on three relevant data sets, and we compare their performance with several recommendation systems recently proposed in the literature. We show that the proposed similarity measures allow to attain an improvement of performances of up to 20\% with respect to existing non-parametric methods, and that the accuracy of a recommendation can vary widely from one specific bipartite network to another, which suggests that a careful choice of the most suitable method is highly relevant for an effective recommendation on a given system. Finally, we studied how an increasing presence of random links in the network affects the recommendation scores, and we found that one of the two recommendation algorithms introduced here can systematically outperform the others in noisy data sets.