A Two Step Approach to Weighted Bipartite Link Recommendations

TL;DR

This paper introduces a two-step bipartite graph-based algorithm for link recommendation that considers frequency and similarity, demonstrating improved accuracy over baseline methods on Epinions and Movielens datasets.

Contribution

It proposes a novel two-step approach leveraging bipartite graphs for link recommendation, incorporating frequency and similarity measures to enhance prediction accuracy.

Findings

Achieved roughly 14% error rate on datasets

Improved upon baseline recommendation methods

Potential for further refinement and accuracy enhancement

Abstract

Many real world person-person or person-product relationships can be modeled graphically. More specifically, bipartite graphs can be especially useful when modeling scenarios that involve two disjoint groups. As a result, many existing papers have utilized bipartite graphs for the classical link recommendation problem. In this paper, using the principle of bipartite graphs, we present another approach to this problem with a two step algorithm that takes into account frequency and similarity between common edges to make recommendations. We test this approach with bipartite data gathered from the Epinions and Movielens data sources, and find it to perform with roughly 14 percent error, which improves upon baseline results. This is a promising result, and can be refined to generate even more accurate recommendations.