Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System

Abstract: The general aim of the recommender system is to provide personalized suggestions to users, which is opposed to suggesting popular items. However, the normal training paradigm, i.e., fitting a recommender model to recover the user behavior data with pointwise or pairwise loss, makes the model biased towards popular items. This results in the terrible Matthew effect, making popular items be more frequently recommended and become even more popular. Existing work addresses this issue with Inverse Propensity Weighting (IPW), which decreases the impact of popular items on the training and increases the impact of long-tail items. Although theoretically sound, IPW methods are highly sensitive to the weighting strategy, which is notoriously difficult to tune. In this work, we explore the popularity bias issue from a novel and fundamental perspective -- cause-effect. We identify that popularity bias lies in the direct effect from the item node to the ranking score, such that an item's intrinsic property is the cause of mistakenly assigning it a higher ranking score. To eliminate popularity bias, it is essential to answer the counterfactual question that what the ranking score would be if the model only uses item property. To this end, we formulate a causal graph to describe the important cause-effect relations in the recommendation process. During training, we perform multi-task learning to achieve the contribution of each cause; during testing, we perform counterfactual inference to remove the effect of item popularity. Remarkably, our solution amends the learning process of recommendation which is agnostic to a wide range of models -- it can be easily implemented in existing methods. We demonstrate it on Matrix Factorization (MF) and LightGCN [20]. Experiments on five real-world datasets demonstrate the effectiveness of our method.

• The paper presents a novel causal framework that uses counterfactual reasoning to remove the direct effects of item popularity from recommendations.
• It employs a multi-task learning approach to decompose user-item interactions into key factors like popularity, matching, and conformity.
• Empirical evaluations on MF and LightGCN models across five datasets demonstrate improved accuracy and increased long-tail item coverage.

An Overview of Model-Agnostic Counterfactual Reasoning for Mitigating Popularity Bias in Recommender Systems

This paper addresses a fundamental and persistent issue in recommender systems: the prevalence of popularity bias. Popularity bias arises when commonly interacted-with items overshadow less popular ones in recommendations. This phenomenon detracts from the system's ability to truly learn and reflect user-item preference alignments. To tackle this challenge, the authors introduce a model-agnostic framework that utilizes counterfactual reasoning to mitigate the direct effects of item popularity, thus aligning the recommendations more closely with user preferences.

The paper begins by outlining the prevalence of popularity bias in recommender systems, attributing this to the imbalanced distribution of user interactions, typically characterized by a long-tail pattern where a few items receive most of the interactions. Existing approaches attempt to counteract this bias by reducing the influence of frequently interacted-with items during training; however, these methods often lack a nuanced approach to addressing individual user cases, resulting in limited success.

The authors propose a novel perspective by analyzing the recommendation task through a causal lens. They construct a causal graph that considers three primary factors influencing user-item interactions: user-item matching, item popularity, and user conformity. The framework, termed Model-Agnostic Counterfactual Reasoning (MACR), leverages this causal framework to disentangle these influences.

MACR operates by reformulating the recommendation inference process to remove the direct effect imparted by item popularity. During the training phase, the algorithm employs a multi-task learning strategy to decompose user-item interactions into contributions from each cause. By performing counterfactual inference, MACR enables the prediction process to simulate a world devoid of item popularity's direct effect, thereby enhancing the model's ability to focus on genuine user-item matches.

The authors demonstrate the applicability of MACR to a broad range of existing recommendation models with minimal modifications. Implementations on Matrix Factorization (MF) and LightGCN reinforce its flexibility and effectiveness, with empirical results showing significant improvements across five datasets—even against a backdrop where traditional methods have struggled.

Significantly, MACR shows substantial performance gains over prior methods like ExpoMF, DICE, and regularization-based approaches. By systematically addressing the direct effects of item popularity while remaining model-agnostic, MACR not only enhances recommendation accuracy but also improves coverage of long-tail items, increasing diversity without sacrificing user relevance.

Looking ahead, this work opens several avenues for future exploration. Further investigations might explore the integration of side information or the joint elimination of other biases within recommender systems. Additionally, extending the framework’s principles beyond recommendation tasks could yield exciting applications in related domains impacted by skewed data distributions.

In conclusion, the authors present a robust framework that challenges conventional training paradigms in recommender systems by incorporating causal reasoning. The MACR framework advances the field by offering a theoretically sound and practically effective approach to mitigating popularity bias, promoting a more equitable distribution of recommendations, and maximizing user satisfaction.