Causal Intervention for Leveraging Popularity Bias in Recommendation (2105.06067v1)

Abstract: Recommender system usually faces popularity bias issues: from the data perspective, items exhibit uneven (long-tail) distribution on the interaction frequency; from the method perspective, collaborative filtering methods are prone to amplify the bias by over-recommending popular items. It is undoubtedly critical to consider popularity bias in recommender systems, and existing work mainly eliminates the bias effect. However, we argue that not all biases in the data are bad -- some items demonstrate higher popularity because of their better intrinsic quality. Blindly pursuing unbiased learning may remove the beneficial patterns in the data, degrading the recommendation accuracy and user satisfaction. This work studies an unexplored problem in recommendation -- how to leverage popularity bias to improve the recommendation accuracy. The key lies in two aspects: how to remove the bad impact of popularity bias during training, and how to inject the desired popularity bias in the inference stage that generates top-K recommendations. This questions the causal mechanism of the recommendation generation process. Along this line, we find that item popularity plays the role of confounder between the exposed items and the observed interactions, causing the bad effect of bias amplification. To achieve our goal, we propose a new training and inference paradigm for recommendation named Popularity-bias Deconfounding and Adjusting (PDA). It removes the confounding popularity bias in model training and adjusts the recommendation score with desired popularity bias via causal intervention. We demonstrate the new paradigm on latent factor model and perform extensive experiments on three real-world datasets. Empirical studies validate that the deconfounded training is helpful to discover user real interests and the inference adjustment with popularity bias could further improve the recommendation accuracy.

An Essay on "Causal Intervention for Leveraging Popularity Bias in Recommendation"

The paper "Causal Intervention for Leveraging Popularity Bias in Recommendation" addresses the perennial challenge of popularity bias in recommender systems. It navigates through the nuanced landscape of biases, distinguishing between detrimental biases that degrade recommendation performance and those that offer potential benefits. The authors propose a refined methodology that not only mitigates the adverse effects of popularity bias but also intelligently leverages the beneficial aspects to enhance recommendation accuracy.

Overview of the Methodology

The paper introduces a novel framework entitled Popularity-bias Deconfounding and Adjusting (PDA). At its core, the framework utilizes causal graphs to accurately ascertain the role of item popularity as a confounder in the data generation process for recommendations. By employing do-calculus, the authors isolate the negative effects of popularity bias, leading to what they term "deconfounded training." This methodology discards spurious correlations induced by popularity bias and aims to retain only intrinsic user-item preference signals.

The design of PDA is further solidified with an inference adjustment mechanism that allows the introduction of desired popularity bias. This is operationalized through forecasted popularity, which is injected during the inference stage to capitalize on items' potential for increased popularity. This dual approach—mitigating negative biases while leveraging positive trends—sets the framework apart from existing methodologies which often blanket-eliminate popularity bias without differentiation.

Evaluation and Results

Empirical evaluations are conducted across three real-world datasets: Kwai, Douban, and Tencent. The authors measure performance using standard metrics such as Recall, Precision, Hit Ratio (HR), and Normalized Discounted Cumulative Gain (NDCG). PDA consistently outperforms baseline models, demonstrating significant performance improvements. For instance, on the Kwai dataset, PD—the foundational component of PDA—achieves an impressive improvement of approximately 241% in Hit Ratio over BPRMF, a traditional collaborative filtering method.

The authors also conduct a stratified analysis to juxtapose the recommendation rates of items differing in popularity. This analysis reveals that PDA maintains a more equitable distribution of recommendation opportunities across the popularity spectrum, in contrast to baseline methods that exhibit a tendency to amplify bias towards popular items.

Theoretical and Practical Implications

The theoretical implications of this work are multifaceted. It provides a robust framework for employing causal inference techniques within recommendation systems, guiding future research to consider data generation processes beyond traditional correlational models. Practically, the PDA framework can significantly impact how recommendation systems are deployed in environments where item popularity fluctuates rapidly, such as e-commerce platforms or content-sharing services.

Future Directions

The authors acknowledge that their approach can be extended in multiple dimensions. Future research could optimize the popularity prediction mechanism at the core of PDA, integrate graph-based representation learning, or expand the framework to concurrently address multiple biases (e.g., positional, temporal). As the field of causal inference in recommendation evolves, these extensions could collectively lead to systems that consistently deliver both fairness and accuracy.

In conclusion, the paper offers a compelling case for re-evaluating the management of popularity bias in recommendation systems. By demarcating and utilizing the beneficial patterns in data, the proposed PDA framework presents a nuanced approach that has both immediate utility and long-term promise in advancing the quality of recommendations.