Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Disentangling Popularity and Quality: An Edge Classification Approach for Fair Recommendation

Nemat Gholinejad, Mostafa Haghir Chehreghani

TL;DR

The paper tackles popularity bias in graph-based recommender systems by reframing the problem as edge classification to separate quality-driven relevance from popularity-driven exposure. It replaces the traditional $BPR$ objective with cross-entropy in a cost-sensitive framework, introducing a disentangled popularity-and-quality mechanism and a two-component algorithm. Empirical results on multiple real-world datasets show substantial fairness improvements (approximately $2\%$ to $74\%$) while preserving competitive accuracy, with ablations confirming the importance of the fairness-oriented components. This approach enables fairer recommendations by prioritizing genuinely relevant, albeit less popular, items without sacrificing user satisfaction or overall performance.

Abstract

Graph neural networks (GNNs) have proven to be an effective tool for enhancing the performance of recommender systems. However, these systems often suffer from popularity bias, leading to an unfair advantage for frequently interacted items, while overlooking high-quality but less popular items. In this paper, we propose a GNN-based recommendation model that disentangles popularity and quality to address this issue. Our approach introduces an edge classification technique to differentiate between popularity bias and genuine quality disparities among items. Furthermore, it uses cost-sensitive learning to adjust the misclassification penalties, ensuring that underrepresented yet relevant items are not unfairly disregarded. Experimental results demonstrate improvements in fairness metrics by approximately 2-74%, while maintaining competitive accuracy, with only minor variations compared to state-of-the-art methods.

Disentangling Popularity and Quality: An Edge Classification Approach for Fair Recommendation

TL;DR

The paper tackles popularity bias in graph-based recommender systems by reframing the problem as edge classification to separate quality-driven relevance from popularity-driven exposure. It replaces the traditional objective with cross-entropy in a cost-sensitive framework, introducing a disentangled popularity-and-quality mechanism and a two-component algorithm. Empirical results on multiple real-world datasets show substantial fairness improvements (approximately to ) while preserving competitive accuracy, with ablations confirming the importance of the fairness-oriented components. This approach enables fairer recommendations by prioritizing genuinely relevant, albeit less popular, items without sacrificing user satisfaction or overall performance.

Abstract

Graph neural networks (GNNs) have proven to be an effective tool for enhancing the performance of recommender systems. However, these systems often suffer from popularity bias, leading to an unfair advantage for frequently interacted items, while overlooking high-quality but less popular items. In this paper, we propose a GNN-based recommendation model that disentangles popularity and quality to address this issue. Our approach introduces an edge classification technique to differentiate between popularity bias and genuine quality disparities among items. Furthermore, it uses cost-sensitive learning to adjust the misclassification penalties, ensuring that underrepresented yet relevant items are not unfairly disregarded. Experimental results demonstrate improvements in fairness metrics by approximately 2-74%, while maintaining competitive accuracy, with only minor variations compared to state-of-the-art methods.

Paper Structure

This paper contains 29 sections, 12 equations, 7 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related work
  3. GNN-based recommendation
  4. Data-driven fair recommendation
  5. Popularity debiasing
  6. Preliminaries
  7. Notations and definitions
  8. Long-tail distribution
  9. Light graph convolutional model
  10. Our proposed method
  11. Disentangled popularity and quality
  12. Edge classification
  13. Cost-sensitive learning
  14. The algorithm
  15. Disentangled popularity and quality
  16. ...and 14 more sections

Figures (7)

  • Figure 1: Degree distribution of items in the Bookcrossing dataset.
  • Figure 2: Studying the effect of $\lambda$, over the Electronics dataset, while setting $\gamma$ to $20$.
  • Figure 3: Studying the effect of $\lambda$, over the bookcrossing dataset, while setting $\gamma$ to $20$.
  • Figure 4: Studying the effect of $\lambda$, over the CDs dataset, while setting $\gamma$ to $20$.
  • Figure 5: Studying the effect of $\gamma$ over the Electronics dataset.
  • ...and 2 more figures

Theorems & Definitions (1)

  • Definition 1