Recommendation Algorithms to Increase Equitable Access to Influencers in a Network

TL;DR

This paper introduces new recommendation algorithms aimed at increasing fairness in networks by improving access to influencers, validated through extensive simulations on real-world data sets.

Contribution

It develops a novel fairness measure and a recommendation algorithm that balances triangle closeness and diversity, demonstrating improved fairness over existing methods.

Findings

The proposed algorithm achieves the best fairness in simulations.

The algorithm is robust to different parameter choices.

Insights are provided on when to use different importance sampling methods.

Abstract

We propose novel recommendation algorithms to improve fairness in networks. Fairness is measured by how close different nodes are to influencers in the network. To allow for easy comparison of fairness across graphs of different sizes, our fairness measure is normalized to the same measure on a synthetic power-law graph of the same size. We experimented with the Erdos-Renyi and Barabasi-Albert graphs and found the latter to be more robust in terms of normalization. In addition to developing a new fairness measure, we propose a new node recommendation algorithm to increase fairness in networks. Our algorithm works by recommending a target node based on the number of triangles between the source and target node with probability P, and with probability 1-P, it introduces weak ties and diversity in the network by recommending nodes using an importance sampling algorithm. This sampling algorithm is based on a polynomial function of the degree of the target node and its distance from the influencer set. Through extensive simulations on three real-world network data sets and comparing seven different algorithms, we show that the algorithm which recommends target nodes with probability proportional to the square of the ratio of the degree of the target node to distance to influencer achieves the best fairness. We show the robustness of the algorithm to different parameter choices and provide insights on when to use the different importance sampling methods based on the structure of the network. We also provide a generalization of our method for disconnected graphs.