Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

General Debiasing for Graph-based Collaborative Filtering via Adversarial Graph Dropout

An Zhang, Wenchang Ma, Pengbo Wei, Leheng Sheng, Xiang Wang

TL;DR

This work tackles bias amplification in graph-based collaborative filtering by introducing AdvDrop, an adversarial graph dropout framework. It learns bias-mitigated and bias-aware subgraphs via a learnable bias measure $P_B$ and enforces invariant representations across views through a contrastive objective, while jointly optimizing with a debiased recommendation loss. An adversarial bias identification stage iteratively discovers bias distributions, enabling generalized debiasing across both general distribution shifts and specific biases like popularity and attribute fairness. Experiments across five public datasets demonstrate consistent improvements in ranking metrics and reduced prediction bias, highlighting AdvDrop's potential for robust, out-of-distribution generalization in graph-based recommender systems.

Abstract

Graph neural networks (GNNs) have shown impressive performance in recommender systems, particularly in collaborative filtering (CF). The key lies in aggregating neighborhood information on a user-item interaction graph to enhance user/item representations. However, we have discovered that this aggregation mechanism comes with a drawback, which amplifies biases present in the interaction graph. For instance, a user's interactions with items can be driven by both unbiased true interest and various biased factors like item popularity or exposure. However, the current aggregation approach combines all information, both biased and unbiased, leading to biased representation learning. Consequently, graph-based recommenders can learn distorted views of users/items, hindering the modeling of their true preferences and generalizations. To address this issue, we introduce a novel framework called Adversarial Graph Dropout (AdvDrop). It differentiates between unbiased and biased interactions, enabling unbiased representation learning. For each user/item, AdvDrop employs adversarial learning to split the neighborhood into two views: one with bias-mitigated interactions and the other with bias-aware interactions. After view-specific aggregation, AdvDrop ensures that the bias-mitigated and bias-aware representations remain invariant, shielding them from the influence of bias. We validate AdvDrop's effectiveness on five public datasets that cover both general and specific biases, demonstrating significant improvements. Furthermore, our method exhibits meaningful separation of subgraphs and achieves unbiased representations for graph-based CF models, as revealed by in-depth analysis. Our code is publicly available at https://github.com/Arthurma71/AdvDrop.

General Debiasing for Graph-based Collaborative Filtering via Adversarial Graph Dropout

TL;DR

This work tackles bias amplification in graph-based collaborative filtering by introducing AdvDrop, an adversarial graph dropout framework. It learns bias-mitigated and bias-aware subgraphs via a learnable bias measure and enforces invariant representations across views through a contrastive objective, while jointly optimizing with a debiased recommendation loss. An adversarial bias identification stage iteratively discovers bias distributions, enabling generalized debiasing across both general distribution shifts and specific biases like popularity and attribute fairness. Experiments across five public datasets demonstrate consistent improvements in ranking metrics and reduced prediction bias, highlighting AdvDrop's potential for robust, out-of-distribution generalization in graph-based recommender systems.

Abstract

Graph neural networks (GNNs) have shown impressive performance in recommender systems, particularly in collaborative filtering (CF). The key lies in aggregating neighborhood information on a user-item interaction graph to enhance user/item representations. However, we have discovered that this aggregation mechanism comes with a drawback, which amplifies biases present in the interaction graph. For instance, a user's interactions with items can be driven by both unbiased true interest and various biased factors like item popularity or exposure. However, the current aggregation approach combines all information, both biased and unbiased, leading to biased representation learning. Consequently, graph-based recommenders can learn distorted views of users/items, hindering the modeling of their true preferences and generalizations. To address this issue, we introduce a novel framework called Adversarial Graph Dropout (AdvDrop). It differentiates between unbiased and biased interactions, enabling unbiased representation learning. For each user/item, AdvDrop employs adversarial learning to split the neighborhood into two views: one with bias-mitigated interactions and the other with bias-aware interactions. After view-specific aggregation, AdvDrop ensures that the bias-mitigated and bias-aware representations remain invariant, shielding them from the influence of bias. We validate AdvDrop's effectiveness on five public datasets that cover both general and specific biases, demonstrating significant improvements. Furthermore, our method exhibits meaningful separation of subgraphs and achieves unbiased representations for graph-based CF models, as revealed by in-depth analysis. Our code is publicly available at https://github.com/Arthurma71/AdvDrop.
Paper Structure (31 sections, 2 theorems, 22 equations, 5 figures, 9 tables, 1 algorithm)

This paper contains 31 sections, 2 theorems, 22 equations, 5 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Collaborative Filtering (CF)
  4. Graph-based CF Framework
  5. Methodology
  6. Debiased Representation Learning
  7. Bias Identification
  8. Model Optimization for AdvDrop
  9. Experiments
  10. Experimental Settings
  11. Performance w.r.t. General Bias (RQ1)
  12. Performance on Specific Bias (RQ2)
  13. Popularity Bias
  14. Attribute Unfairness
  15. Study of AdvDrop (RQ3)
  16. ...and 16 more sections

Key Result

theorem 1

For a vector of $N$ binary random variables $\mathbf{x}= (x_1,\dots,x_N)^T$, and any function $f$, the gradient of with respect to $\boldsymbol{\phi} = (\phi_1,\dots,\phi_N)^T$, the logits of the Bernoulli probability parameters, can be expressed as: where $\mathbb{I}[\boldsymbol{v}>\sigma(-\phi)]:= (\mathbb{I}[v_1>\sigma(-\phi_1)],\dots,\mathbb{I}[v_N>\sigma(-\phi_N)])^T$, and $\sigma(\cdot)$ i

Figures (5)

  • Figure 1: T-SNE t-sne visualizations of user and item representations learned by MF BPR, LightGCN LightGCN, and our proposed AdvDrop. Note that MF, LightGCN-2, and LightGCN-4 are specialized with zero, two, and four graph convolutional layers, respectively. Subfigures \ref{['fig:gender_mf']}-\ref{['fig:gender_AdvDrop']} show the representation distribution w.r.t. two groups of user gender (i.e., female, male), while Subfigures \ref{['fig:pop_mf']}-\ref{['fig:pop_AdvDrop']} depict the representation distribution w.r.t. three groups of item popularity (i.e., head, middle, tail).
  • Figure 2: Recommendation Performance
  • Figure 3: The overall framework of AdvDrop.
  • Figure 4: Visualization of learned bias measurement function $P_B$w.r.t. item popularity.
  • Figure 5: (a) The overall recommendation performance v.s. epochs during training. (b) $\sim$ (e) The debiasing performance w.r.t. epochs on both user and item attributes.

Theorems & Definitions (2)

  • theorem 1
  • corollary 1