Causal Intervention for Leveraging Popularity Bias in Recommendation

Yang Zhang; Fuli Feng; Xiangnan He; Tianxin Wei; Chonggang Song; Guohui Ling; Yongdong Zhang

Source-linked AI summary

Causal Intervention for Leveraging Popularity Bias in Recommendation

Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, Yongdong Zhang

arXiv:2105.06067v1May 13, 2021cs.IR

TL;DR

Popularity bias creates long-tail exposure and can be amplified by recommender models, yet some popularity reflects item quality or trends. This paper proposes Popularity-bias Deconfounding and Adjusting (PDA), using causal analysis to remove confounding bias during training and inject desired popularity during inference. Experiments on three real-world datasets show that deconfounding and popularity adjustment improve recommendation performance.

Problem
Popularity bias causes uneven item exposure and amplification of popular items, while eliminating it wholesale may discard signals of quality or trends.
Method
PDA models item popularity as a confounder, applies do-calculus for deconfounded training, and causally adjusts recommendation scores during inference.
Results
PDA improves recommendation performance across three real-world datasets, with more than 241% improvement over BPRMF on all Kwai metrics.
Takeaways & Limitations
Popularity bias can be decoupled and selectively leveraged rather than uniformly eliminated to improve recommendation accuracy.
Takeaways & Limitations
The approach is demonstrated on a latent factor model, with extensions to graph-based models and content features left for future work.

Abstract

from arXiv · show

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.

1 INTRODUCTION

Popularity bias creates long-tail exposure patterns and can be amplified by recommendation models, but some popularity reflects item quality or trends. The paper therefore models popularity causally, removes its confounding effect during training, and injects desired popularity during inference through PDA.

Popularity-bias problem: Popularity bias appears as long-tail interaction data and causes recommender models to over-recommend popular items.A few head items receive most interactions, while model training on such data can amplify their dominance in top recommendations.
Existing approaches: Existing bias-aware methods mainly pursue unbiased learning or ranking adjustment, while intentionally boosting less popular items lacks theoretical foundations.The cited approaches include inverse propensity scoring and causal embedding, with practical or methodological limitations.
Motivation: Some popularity reflects intrinsic quality or current trends, so eliminating all popularity signals can suppress items that deserve more recommendations.The paper also considers platforms that want to promote items with potential future popularity.
Causal analysis: Item popularity Z is a common cause of exposure I and interaction probability C, creating a confounding path that amplifies popularity beyond user-interest matching.Popularity affects interactions through conformity and affects exposure because recommendation systems expose popular items more frequently.
PDA framework: PDA uses do-calculus to estimate P(C|do(U, I)) during training and intervenes on popularity through P(C|do(U, I),do(Z)) during inference.The training intervention cuts the popularity-to-exposure path, while inference can impose a desired bias such as forecasted popularity.
Contributions: The framework combines causal analysis, deconfounded training, and popularity intervention, with experiments on three real-world datasets validating its effectiveness.These components are summarized as the paper’s main contributions.

2 PRIMARY KNOWLEDGE

The paper treats recommendation as prediction over sequentially collected stages and studies popularity as a dynamic distributional factor. It defines Drift of Popularity to measure changes between stages and uses the observed drift to motivate forecasting and future-stage adjustment.

Sequential recommendation setting: Historical interaction data is collected across T stages, and the recommender aims to capture preferences for accurate recommendations on stage T+1.The setup distinguishes users and items and focuses on item popularity as a factor affecting next-stage performance.
Popularity representation: Local popularity is based on an item's interaction frequency within a stage and is considered more influential than global popularity.The paper attributes this to periodic retraining and the stronger effect of recent data on exposure and user decisions.
Popularity drift: Drift of Popularity (DP) measures popularity change between two stages using Jensen-Shannon Divergence over item-frequency distributions.Its range is [0, log(2)], and higher values indicate larger popularity drift.
Empirical drift patterns: Popularity drift exists across the three real-world datasets, differs by dataset, and accumulates as the interval between stages increases.Figure 2 compares successive stages through DP(t,t+1) and the first and present stages through DP(1,t).
Implication: Because future popularity differs from past popularity, forecasting its trend and injecting it into recommendations is presented as a viable route to next-stage accuracy.The proposed target is high accuracy on the next-stage data D_T+1.

3 METHODOLOGY

The methodology models item popularity as a confounder that amplifies exposure bias, then removes this effect during training and adjusts popularity during inference. PDA uses causal intervention and a decoupled user-item matching component to support both stages.

3.1 Causal View of Recommendation: Item popularity Z confounds exposed items I and observed interactions C through paths Z→C and Z→I→C.The second path increases exposure to popular items and causes bias amplification.
3.1 Causal View of Recommendation: Popularity also changes learned item embeddings, increasing popular items’ vector lengths and their scores across users.This effect undermines reliable estimation of real user interest.
3.2 Deconfounded Training: Do-calculus estimates P(C|do(U, I)) by cutting the popularity-to-exposure path without requiring interventional experiments.The derivation applies the backdoor criterion, Bayes’ theorem, and independence in the intervened graph.
3.2 Deconfounded Training: The conditional model decouples user-item matching from popularity, enabling extension to collaborative filtering models and fast inference-stage adjustment.The framework uses Matrix Factorization here, while γ controls the strength of the conformity effect and ELU′ ensures positivity.
3.2 Deconfounded Training: The model fits historical interactions with popularity-conditioned probabilities but uses ELU′(f_Θ(u,i)) for deconfounded ranking.This procedure is named Popularity-bias Deconfounding (PD).
3.3 Adjusting Popularity Bias in Inference: PDA combines deconfounded training with inference-stage popularity adjustment, while direct random exposure remains costly, restricted, and potentially harmful to user experience.The inference adjustment uses a forecast popularity value whose drift strength is controlled by α.

4 EXPERIMENTS

The experiments examine whether PD removes popularity bias and whether PDA can inject desired popularity bias to improve recommendation performance. The supplied experiment passage states these as the study’s two research questions.

Research Questions: The experiments ask whether PD removes popularity bias and how it compares with existing methods.This is RQ1.
Research Questions: The experiments ask whether PDA effectively injects desired popularity bias and how much it enhances recommendation performance.This is RQ2.

4.1 Experimental Settings

Experiments evaluate PDA on three real-world datasets using personalization and popularity-aware baselines under common training and ranking protocols.

Datasets: Experiments use Kwai, Douban Movie, and Tencent datasets containing click, rating, and like interactions, respectively.Kwai contains 7,658,510 interactions; Douban Movie contains 7,174,218; Tencent contains 1,816,046.
Baselines: Baseline comparisons include MostPop, BPRMF, xQuAD, BPR-PC, and DICE for popularity-bias-aware recommendation.MostPop ignores personalization, while xQuAD and BPR-PC adjust rankings and DICE learns causal embeddings.
Popularity-aware baselines: Popularity-aware comparisons include MostRecent, BPRMF(t)-pop, and inference-adjusted variants BPRMF-A and DICE-A.These methods introduce or model popularity during recommendation inference for comparison with PDA.
Evaluation protocol: All methods use BPR loss, Adam optimization, validation tuning, L2 regularization search, and early stopping based on Recall@20.The experiments report results at K=20 and K=50 using HR, Recall, Precision, and NDCG under the all-ranking protocol.

4.2 Deconfounding Performance (RQ1)

PDA’s deconfounded training improves top-K recommendation performance and produces recommendation rates closer to the training distribution. The analysis also shows that local popularity computation is important, while manually adjusting unpopular-item exposure does not directly establish user interest.

Overall Performance: PD achieves the best performance and consistently outperforms all baselines across the three datasets.Relative improvement over corresponding baselines is at least 119% on Kwai, 34% on Douban, and 24% on Tencent.
Overall Performance: Higher popularity drift corresponds to larger advantages for PD over conventional formulations in the reported datasets.The paper attributes the largest improvement on Kwai to its substantially larger popularity drift.
Recommendation Analysis: BPRMF amplifies popularity bias, with recommendation rates declining almost linearly from popular group 1 to unpopular group 10.Figure 4 compares recommendation rates across item-popularity groups, with panels for Kwai, Douban, and Tencent.
Recommendation Analysis: PD produces flatter recommendation-rate curves that more closely resemble the uniform rates of the training set.Recommendation rate divides recommendations from each item group by the total number of recommendations; lower variation indicates greater uniformity.
Evaluation scope: The analysis omits ranking-adjustment recommendation rates because those methods manually increase recommendations for unpopular items.The evaluation instead emphasizes whether learned embeddings reveal users’ real interests.
Local popularity: PD-G performs far worse than PD on Recall@50 and NDCG@50, verifying the importance of calculating item popularity locally.PD-G computes popularity over the entire training set rather than at local training stages.

4.3 Performance of Adjusting Popularity (RQ2)

PDA improves recommendation by injecting desired popularity bias at inference, outperforming popularity-aware baselines and benefiting from more precise popularity predictions.

Comparisons with Baselines: Introducing desired popularity bias into recommendations improves performance compared with models without inference adjustment.The comparison includes PDA, DICE-A, BPRMF-A, and MostRecent against PD, DICE, BPRMF, and MostPop.
Comparisons with Baselines: PDA outperforms BPRMF(t)-pop, BPRMF-A, and DICE-A by deconfounding training and intervening during inference.The authors attribute this advantage to preventing popularity amplification while injecting desired popularity bias at recommendation time.
More Refined Predictions for Popularity: More precise popularity prediction further improves PDA performance across the evaluated settings.The experiment splits the last training stage into N sub-stages and evaluates N = 1, 2, 4, 8, 16.
Comparisons with Baselines: 241%+ relative improvement is achieved by PDA over BPRMF on all metrics for Kwai.On Douban and Tencent, PDA improves recall and NDCG by at least 97% and precision and HR by at least 57%.

5 RELATED WORKS

Related work addresses popularity bias through diagnosis, debiasing, popularity utilization, and causal recommendation, while PDA combines deconfounding with selective bias use.

Popularity Bias: Popularity-bias research studies its prevalence, effects across recommendation models, and relationships with fairness and calibration.Prior work reports larger popularity effects for BPR-based models than point-wise models.
Popularity Utilization: Other recommendation methods use temporal or local popularity, but PDA distinguishes popularity bias to remove from popularity drift to leverage.Prior examples target music, news, or general recommendation settings using drifted popularity.
Causal Recommendation: Causal recommendation work includes confounding-effect methods that estimate causal effects or learn real interests, alongside counterfactual methods that model unobserved treatments.These approaches use exposure data, reweighting, quasi-experimental frameworks, or limited unbiased data.

6 CONCLUSION

The paper frames popularity as a confounder and proposes PDA to deconfound training while leveraging selected popularity during recommendation. Experiments on three real-world datasets support this framework, while future work targets broader bias factors and model classes.

Conclusion: Item popularity confounds estimation of P(C|U, I), so some popularity bias should be decoupled rather than uniformly eliminated.The paper argues that selective use of popularity can improve recommendation.
Conclusion: PDA combines deconfounded training with popularity-bias adjustment to leverage popularity for recommendation.The framework is evaluated through experiments on three real-world datasets.
Conclusion: The paper identifies pure data-driven recommendation models as limited and expresses the causal data-generation process through a causal graph.The authors propose extending the paradigm to position bias, selection bias, graph-based models, and content features.