Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity

Shiyin Tan, Dongyuan Li, Renhe Jiang, Zhen Wang, Xingtong Yu, Manabu Okumura

TL;DR

The paper addresses popularity bias in recommender systems by acknowledging that users have evolving preferences for item popularity. It introduces CausalEPP, a method that defines evolving personal popularity, builds a causal graph that disentangles item quality from popularity, and employs deconfounded training to estimate causal effects. Temporal evolution is incorporated through moving-average-based interventions during inference, enabling forecast-guided adjustments to conformality. Empirical results across three benchmarks show state-of-the-art performance and meaningful debiasing gains, with ablations validating the contribution of evolving popularity, quality disentangling, and temporal intervention to improved recommendations.

Abstract

Popularity bias occurs when popular items are recommended far more frequently than they should be, negatively impacting both user experience and recommendation accuracy. Existing debiasing methods mitigate popularity bias often uniformly across all users and only partially consider the time evolution of users or items. However, users have different levels of preference for item popularity, and this preference is evolving over time. To address these issues, we propose a novel method called CausalEPP (Causal Intervention on Evolving Personal Popularity) for taming recommendation bias, which accounts for the evolving personal popularity of users. Specifically, we first introduce a metric called {Evolving Personal Popularity} to quantify each user's preference for popular items. Then, we design a causal graph that integrates evolving personal popularity into the conformity effect, and apply deconfounded training to mitigate the popularity bias of the causal graph. During inference, we consider the evolution consistency between users and items to achieve a better recommendation. Empirical studies demonstrate that CausalEPP outperforms baseline methods in reducing popularity bias while improving recommendation accuracy.

Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity

TL;DR

The paper addresses popularity bias in recommender systems by acknowledging that users have evolving preferences for item popularity. It introduces CausalEPP, a method that defines evolving personal popularity, builds a causal graph that disentangles item quality from popularity, and employs deconfounded training to estimate causal effects. Temporal evolution is incorporated through moving-average-based interventions during inference, enabling forecast-guided adjustments to conformality. Empirical results across three benchmarks show state-of-the-art performance and meaningful debiasing gains, with ablations validating the contribution of evolving popularity, quality disentangling, and temporal intervention to improved recommendations.

Abstract

Popularity bias occurs when popular items are recommended far more frequently than they should be, negatively impacting both user experience and recommendation accuracy. Existing debiasing methods mitigate popularity bias often uniformly across all users and only partially consider the time evolution of users or items. However, users have different levels of preference for item popularity, and this preference is evolving over time. To address these issues, we propose a novel method called CausalEPP (Causal Intervention on Evolving Personal Popularity) for taming recommendation bias, which accounts for the evolving personal popularity of users. Specifically, we first introduce a metric called {Evolving Personal Popularity} to quantify each user's preference for popular items. Then, we design a causal graph that integrates evolving personal popularity into the conformity effect, and apply deconfounded training to mitigate the popularity bias of the causal graph. During inference, we consider the evolution consistency between users and items to achieve a better recommendation. Empirical studies demonstrate that CausalEPP outperforms baseline methods in reducing popularity bias while improving recommendation accuracy.

Paper Structure

This paper contains 18 sections, 1 theorem, 14 equations, 7 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Popularity Bias in Recommendation System
  4. Debiasing Recommendation
  5. Comparison with Current algorithms
  6. Preliminaries
  7. Methodology
  8. Evolving Personal Popularity
  9. Causal Graph
  10. Deconfounded Training
  11. Intervened Inference
  12. Experiment
  13. Experimental Settings
  14. Performance Comparison (RQ1)
  15. Ablation Study (RQ2)
  16. ...and 3 more sections

Key Result

Theorem 1

If a random variable $X$ with the probability distribution $P(X)$ has the expectation $\mathbb{E}(X)$, given the non-linear function $f: \mathcal{G} \to \mathbb{R}$ where $\mathcal{G}$ is a closed subset of $\mathbb{R}$, following: then the inequality holds: $|\mathbb{E}[f(X)] - f(\mathbb{E}(X))| \leq T (\rho_\beta^\beta + \rho_\gamma^\gamma),$ where $\rho_\beta = \sqrt[\beta]{\mathbb{E}[|X - \m

Figures (7)

  • Figure 1: (A) Without popularity debiasing, recommender systems recommend popular items to all users. Current work mitigates popularity bias uniformly across all users and ignores the user's preference toward popular items. Our work considers both users' preference for popular items and their hobby. (B) Without the evolution of both users and items, recommender systems will recommend items that consider the overall popularity across all times. With only item evolution, they will recommend currently popular items to all users. Our work considers both user and item evolutions, which will consider consistency between the two evolutions.
  • Figure 2: (A) An example of causal graph with confounder. (B) Deconfounded training by eliminating the influence of variable Z on variable T.
  • Figure 3: (A): Causal graph to describe the recommendation process that incorporates evolving personal popularity "S" into the conformity effect and disentangles quality "Q" from popularity "P". (B): During training, we cut off the influence from local popularity to items for deconfounded training. (C): During inference, we intervene evolving personal popularity "$s^*$" and local popularity "$p^*$" for bias adjustment.
  • Figure 4: We illustrate the performance (%) on the Amazon-Music dataset for both LightGCN and MF backbones with varying consistency ratios $\alpha$. The shaded area highlights the performance gain brought from the intervention.
  • Figure 5: The ablation analysis of quality across global popularity on the Douban-Movie dataset. "w/o quality loss" indicates the quality loss was removed. "original" indicates the ground truth quality (the average rating of each group of items) and "CausalEPP" indicates the quality of CausalEPP.
  • ...and 2 more figures

Theorems & Definitions (4)

  • Definition 1: Local Popularity
  • Definition 2: Moving Average
  • Definition 3: Evolving Personal Popularity
  • Theorem 1: Jensen's inequality of non-linear function