Pith. sign in

REVIEW 5 major objections 5 minor 77 references

Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Popularity-bias debiasing should follow each user's evolving taste for popular items, and CausalEPP does exactly that with causal intervention.

desk verdict A useful, well-tested empirical recipe for personalized popularity debiasing, but the causal grounding rests on an unverified Jensen-gap approximation. read the letter →

arxiv 2505.14310 v2 pith:SBDOD6OX submitted 2025-05-20 cs.IR cs.LG

Shiyin Tan , Dongyuan Li , Renhe Jiang , Zhen Wang , Xingtong Yu , Manabu Okumura This is my paper
classification cs.IRcs.LG
keywords popularitybiasrecommendationdebiasingcausalinferencebackdooradjustmenttemporalevolutionpersonaldeconfoundedtraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that popularity bias in recommender systems is not a single uniform distortion: different users have different, time-varying appetites for popular items, so debiasing should be personalized and should track how both users and items evolve. To that end it introduces CausalEPP, a method built around a metric called evolving personal popularity that measures how often a user clicks popular items in a recent window. The method models the recommendation process with a causal graph, cuts off the confounding path from local popularity to item exposure via backdoor adjustment, separates item quality from popularity, and during inference forecasts future popularity and personal popularity with moving averages to set intervention values. On three datasets and two backbones the paper reports that CausalEPP outperforms ten debiasing baselines, with recall improvements up to 20.4% before intervention, and that it shifts the share of popular-item recommendations toward the ground-truth ratio. The payoff, if the claim holds, is that recommender systems can reduce popularity bias without uniformly penalizing popular items or ignoring users who genuinely prefer them.

What carries the argument

The load-bearing object is the evolving personal popularity $s^t_u$, the fraction of a user's clicks in a recent window that land on items whose local popularity exceeds the top-20% threshold, paired with a causal graph $U\to S\to C\gets P\gets I$ and $Q\to Y$. The deconfounding step is backdoor adjustment: cutting the edge $P\to I$ yields $\mathbb{E}_p[\hat{y}]$, which the paper approximates by evaluating the prediction at the average popularity $\mathbb{E}[p]$ using a Jensen-gap bound (Theorem 1). The conformity effect $c=e^{-\alpha|s-p|}p\,\mathrm{MLP}(i)$ carries the personalization, and the quality loss $\mathcal{L}_Q$ forces item quality $q_i$ to be ordered like global popularity. During inference, moving-average gradients forecast $p^*$ and $s^*$, which are used as intervention values in the same conformity formula.

What would settle it

On a held-out time split, compute the trained model's exact expectation $\mathbb{E}_p[\hat{y}_{ui}]$ by sampling local popularity values from the empirical distribution and compare it with the approximation $\hat{y}(\mathbb{E}[p])$; if the average absolute gap is comparable to the reported performance differences between CausalEPP and the best baseline, the deconfounding claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that popularity bias can be tamed by treating it as a confounder problem in a causal graph whose nodes include user, item, local popularity, evolving personal popularity, item quality, conformity, matching score, and click outcome. The paper derives the backdoor-adjusted distribution $P(Y|U,\mathrm{do}(I=i),Q)=\mathbb{E}_p[P(Y|q,C(S(u),p),M(u,i))]$, models the conformity effect as $c=e^{-\alpha|s-p|}p\,\mathrm{MLP}(i)$, and trains with a BPR loss plus a quality loss that forces learned quality to follow global popularity. Deconfounded training cuts the path $P\to I$, and intervened inference sets $P=p^*$ and $S=s^*$ using moving-average gradients to forecast future popularity and personal preference. The empirical claim is that this combination consistently beats ten state-of-the-art debiasing baselines across Ciao, Amazon-Music, and Douban-Movie on MF and LightGCN backbones, and that it moves the proportion of recommended popular items from 75.8% to 55.5% on Amazon-Music and from 93.9% to 64.1% on Douban-Movie.

Load-bearing premise

The training objective swaps the expectation over local popularity for the prediction at the average popularity, and the argument that this swap is safe rests on a Jensen-gap bound whose size is never computed for the trained network; if that gap is large, the loss being optimized is not the backdoor-adjusted causal objective.

Editorial extensions

If this is right

  • Systems using CausalEPP should recommend popular items mainly to users whose recent history shows a taste for them, and shift away from popular items for users who do not show that taste.
  • The intervention step extrapolates current trends, so recommendations should track seasonal or faddish items more quickly than static debiasing methods do.
  • Separating quality from popularity lets the model keep recommending genuinely good popular items instead of suppressing all popular items uniformly.
  • The reported reductions in popular-item recommendation share would bring recommendation distributions closer to the ground-truth interaction ratios, which is the paper's operational definition of less popularity bias.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The Jensen-gap approximation in Eqs. (8)-(9) means the method's theoretical grounding would be firmer if the gap were computed per dataset; a cheap validation check would be sampling local popularity values and comparing $\mathbb{E}_p[\hat{y}]$ to $\hat{y}(\mathbb{E}[p])$ for the trained model.
  • The same causal recipe could transfer to other settings with time-stamped exposure, such as news feeds or video platforms, where user sensitivity to trending content also drifts; the paper only hints at LLM debiasing as future work.
  • Because $\alpha$ controls sensitivity to the popularity gap and is set to 0.5 for all datasets, per-user or per-item $\alpha$ values learned from validation data might yield larger debiasing gains than the current global setting.
  • The forecast intervention assumes the moving-average gradient direction persists; in domains with abrupt popularity spikes the step sizes $\Delta_T$ would need adapting, which the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes CausalEPP, a popularity-debiasing method for recommender systems that introduces an 'evolving personal popularity' metric per user, models the recommendation process with a causal graph that includes item quality, local popularity, and conformity, and applies backdoor adjustment ('deconfounded training') during training. At inference, the method intervenes on local popularity and personal popularity using moving-average forecasts to capture temporal evolution. The authors evaluate CausalEPP on Ciao, Amazon-Music, and Douban-Movie with MF and LightGCN backbones, reporting gains in Recall@20 and reductions in the share of popular-item recommendations, plus ablations of the consistency score, quality loss, and temporal-evolution intervention.

Significance. If the causal claims were fully supported, the paper would be a useful contribution: it addresses personalized and time-varying popularity bias, a recognized gap in the literature, and its empirical package is substantial, including three datasets, two backbones, ten baselines, ablations, and a debiasing analysis. The authors also release code, which aids reproducibility. The derivation of the backdoor adjustment in Eqs. (5a)-(5d) is standard and clearly presented, and the empirical finding that the intervention reduces over-recommendation of popular items is valuable. However, the central theoretical claim that the training objective is the deconfounded causal objective hinges on an unverified approximation, and the inference-time operation is more accurately described as a heuristic feature substitution than as a rigorously derived counterfactual intervention. The quality-loss design also conflates learned quality with global popularity by construction. These issues affect the paper's core causal framing rather than only its presentation.

major comments (5)
  1. [Section 4.3, Eqs. (8)-(9)] The replacement of E_p[\hat y_{ui}] by the prediction evaluated at E[p_i] is not justified by the cited Jensen bound. Theorem 1 provides an upper bound that depends on rho_beta, gamma, beta, and T, but none of these quantities is computed for the trained network, and no argument shows the gap is small in practice. The function c_ui(p)=exp(-alpha|s_u-p|)p*MLP(i) is non-differentiable at p=s_u and grows linearly in p, so the gap can be substantial when local popularity is dispersed. In addition, Eq. (8) moves the expectation through Tanh and through multiplication by Softplus(m_ui), which is a second unquantified Jensen approximation. If this gap is not negligible, the training loss in Eq. (12) is not the backdoor-adjusted objective in Eq. (5d), and the central claim that CausalEPP performs deconfounded training is unsupported. The authors should either compute the bound, or replace the point-evaluation approximation with a Monte Carlo estimate of E_p[\hat y_{ui}] and check whether the results change materially.
  2. [Section 4.4, Eq. (13)] The inference-time 'intervention' is a feature swap in a trained network, not a do-calculus intervention on the causal model. Equation (13) is asserted as a conditional probability, but no derivation is given for why setting P=p* in the trained scoring function corresponds to P(Y|U,do(I=i),Q,do(P=p*)). The model was trained with E[p] rather than with p as an input during training, so its response to out-of-distribution values p* is unvalidated. To support the causal language, the authors should either make the model's dependence on p explicit and verify its counterfactual behavior (e.g., by training with sampled p values and checking that the learned function varies in the intended way), or reframe this component as a heuristic popularity-adjustment step.
  3. [Section 4.3, Eq. (10)] The quality loss enforces q_i>q_j whenever d_i>d_j, so the learned quality is a monotone function of global popularity by construction. This means the causal separation of Q from P in Fig. 3(A) is not achieved in the estimated model: 'quality' is not identified independently of popularity. Figure 5 confirms that the learned quality tracks global popularity, which is a direct consequence of Eq. (10) rather than evidence that a distinct quality factor has been discovered. The claim of disentangling benign quality from harmful popularity therefore needs additional support, such as showing that q_i carries signal beyond a fixed monotone transform of d_i, or comparing with a model in which quality is learned without the monotonicity constraint.
  4. [Section 5.2, Tables 1 and 3] The statement that CausalEPP 'consistently outperforms all state-of-the-art debiasing baselines' is too strong given the reported numbers. With the MF backbone on Ciao, CausalEPP is worse than PPAC and TIDE on Precision@20 (0.0143 vs. 0.0149 and 0.0148) and NDCG@20 (0.0150 vs. 0.0156 and 0.0154). With the LightGCN backbone on Douban-Movie, CausalEPP is slightly worse than PPAC on Precision@20 (-0.3%). The 'up to 20.4%' improvement in the abstract is specifically the Ciao Recall@20 gain with the MF backbone. The claims should be qualified to recall or average rank, or the paper should report aggregate significance tests over all metrics rather than only per-metric best-baseline comparisons.
  5. [Section 5.1, implementation details] The statistical significance claim marked with daggers in Tables 1 and 3 is not fully documented. The paper states that a t-test with p<0.05 was used, but it does not report the number of random seeds, the variance across runs, or how the test was paired. Without this information, the significance markers cannot be verified. Please add the experimental protocol for the significance tests.
minor comments (5)
  1. [Section 5.2] The phrase 'the proposed invention improves the results' should be 'the proposed intervention improves the results'.
  2. [Theorem 1] The theorem title contains a typo: 'Jensen’s ineqality' should be 'Jensen’s inequality'.
  3. [Section 4.1, Eq. (3)] The notation p_t_i and the threshold p_hat_t are introduced clearly, but the dependence of p_hat_t on the top-20% quantile of local popularity should be stated in the definition as well as in the prose, since the definition currently refers to p_i^t > p_hat^t without specifying how p_hat^t is set.
  4. [Section 5.3, Table 4] The prose says the 'w/o evolution' variant has a performance drop of -4.08%, but the table reports per-dataset drops of -5.94%, -3.10%, and -2.92% for Recall@20 and corresponding NDCG@20 drops. Please clarify whether the text reports an average over datasets and metrics, and if so, state that explicitly.
  5. [Section 4.3, Eq. (12)] The BPR loss in Eq. (12) uses the approximated expectation E_p[\hat y] from Eq. (8), but the text does not discuss how the approximation error propagates into the pairwise ranking loss. A brief sentence noting this limitation would help readers interpret the theoretical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the causal derivation is a standard backdoor adjustment, and the approximations, quality loss, and intervention heuristics are explicit modeling choices rather than reductions to their inputs.

full rationale

The paper's central derivation (Eqs. 4-5) is a textbook backdoor adjustment: it sums over the local popularity distribution and does not define popularity in terms of the predicted click score. The model (Eqs. 6-7) is an instance of the adjusted estimand, not a restatement of it. The approximation in Eqs. (8)-(9) replaces E_p[c_ui] with c_ui evaluated at E[p_i], justified by an external Jensen-gap bound (Theorem 1, cited to [18]); the bound is not numerically evaluated, so the approximation may be loose, but this is an accuracy/robustness concern, not circularity, because the approximated quantity is not defined as the target metric. The quality loss (Eq. 10) explicitly enforces q_i to be ordered by global popularity; this is a stated modeling assumption about 'benign' popularity, not a hidden reintroduction of the evaluation label. The inference intervention (Eqs. 14-15) uses moving-average extrapolation with validation-tuned steps; it is a heuristic intervention-value search, not a fitted predictor of recall. No load-bearing self-citations were found: the only theorem invoked is from an external source, and all empirical comparisons are against external baselines on held-out data. Hence the derivation chain is self-contained and no claim reduces by construction to its inputs.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method's free parameters are all tuned on validation data and directly control the strength of the intervention and the definition of popularity. The causal graph and the expectation approximation are the main structural assumptions. No new physical or conceptual entities are introduced; the latent variables (quality, conformity) are standard in the cited literature.

free parameters (8)
  • consistency ratio alpha = 0.5
    Controls how strongly the conformity score decays with the gap between personal popularity and local popularity (Eq. 7); tuned by grid search on validation.
  • loss balance weight lambda = 0.2
    Balances BPR loss and quality loss in Eq. (11); chosen by hyperparameter sensitivity.
  • sliding window w1 = 6 months
    Time window for computing local popularity and evolving personal popularity (Eqs. 1, 3); tuned.
  • moving average window w2 = 10%
    Window for moving average in Eq. (2); set as 10% of time steps.
  • high-popularity quantile = 20%
    Threshold hat{p}^t that defines popular items in Eq. (3); tuned.
  • item intervention step Delta_T_i = 5
    Length of forecast horizon for local popularity in Eq. (14); tuned on validation.
  • user intervention step Delta_T_u = 10
    Length of forecast horizon for personal popularity in Eq. (15); tuned on validation.
  • hidden dimension = 64
    Embedding dimension for LightGCN/MF backbones; tuned.
assumptions (5)
  • standard math Backdoor adjustment (Pearl 2009) is a valid method to estimate P(Y|do(I=i)) when the graph in Fig 3(A) is correct.
    Used in Eq. (5) to justify removing the P->I edge.
  • domain assumption The causal graph in Fig 3(A) is correct, with no unobserved confounders among U, I, S, P, Q, C, M, Y beyond the modeled edges.
    The entire debiasing argument depends on this graph structure.
  • domain assumption Item quality Q is monotonically related to global popularity d_i, as enforced by Eq. (10); this defines quality operationally.
    The quality loss L_Q forces q_i to be ordered by d_i, so 'quality' is assumed to align with long-run popularity.
  • standard math The Jensen gap bound in Theorem 1 applies to the specific f(x)=Tanh(q+c(x)) and f(x)=c(x) with the stated growth conditions.
    Invoked to justify the approximation E_p[yhat] approximately equals yhat(E[p]).
  • domain assumption Moving-average gradient extrapolation in Eqs. (14)-(15) is a reasonable predictor of near-future popularity and personal popularity.
    The temporal intervention relies on this forecasting heuristic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity." pith.science (2026). https://pith.science/paper/SBDOD6OX

@misc{pith2026250514310,
  author       = {Pith},
  title        = {Pith review of: Taming Recommendation Bias with Causal Intervention on Evolving Personal Popularity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SBDOD6OX}},
  note         = {Machine review of arXiv:2505.14310}
}
read the original 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.

Figures

Figures reproduced from arXiv: 2505.14310 by the authors.

Figure 1
Figure 1. (A) Without popularity debiasing, recommender [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (A) An example of causal graph with confounder. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (A): Causal graph to describe the recommendation [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: We illustrate the performance (%) on the Amazon [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The ablation analysis of quality across global popu [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Parameter sensitive performance of recall@20 on the Amazon-music dataset with LightGCN as a backbone. (A): [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Recommendation frequency of popular (blue) and [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 59 canonical work pages

  1. [1]

    Himan Abdollahpouri. 2020. Popularity bias in recommendation: a multi- stakeholder perspective. Ph. D. Dissertation. University of Colorado at Boulder

    work page 2020

  2. [2]

    Himan Abdollahpouri, Masoud Mansoury, Robin Burke, and Bamshad Mobasher

  3. [3]

    Chris Anderson. 2006. The long tail: Why the future of business is selling more for less. Hyperion

    2006

  4. [4]

    Haoyue Bai, Le Wu, Min Hou, Miaomiao Cai, Zhuangzhuang He, Yuyang Zhou, Richang Hong, and Meng Wang. 2024. Multimodality Invariant Learning for Multimedia-Based New Item Recommendation. In Proceedings of the 47th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, USA, July 14-18, 2024 . AC...

    work page 2024

  5. [5]

    Alejandro Bellogín, Ludovico Boratto, Styliani Kleanthous, Elisabeth Lex, Francesca Maridina Malloci, and Mirko Marras. 2024. International Workshop on Algorithmic Bias in Search and Recommendation (BIAS). In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Washington DC, USA, Jul...

    work page 2024

  6. [6]

    Richard Blundell. 1988. Consumer behaviour: theory and empirical evidence—a survey. The economic journal 98, 389 (1988), 16–65

    work page 1988

  7. [7]

    Miaomiao Cai, Lei Chen, Yifan Wang, Haoyue Bai, Peijie Sun, Le Wu, Min Zhang, and Meng Wang. 2024. Popularity-Aware Alignment and Contrast for Mitigating Popularity Bias. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024 . ACM, 187–198

    work page 2024

  8. [8]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He

Show all 77 references
  1. [9]

    Jiaju Chen, Wenjie Wang, Chongming Gao, Peng Wu, Jianxiong Wei, and Qing- song Hua. 2024. Treatment Effect Estimation for User Interest Exploration on Recommender Systems. In Proceedings of the 47th International ACM SIGIR Confer- ence on Research and Development in Informatio...

    2024

  2. [10]

    Jiajia Chen, Jiancan Wu, Jiawei Chen, Xin Xin, Yong Li, and Xiangnan He. 2024. How graph convolutions amplify popularity bias for recommendation? Frontiers Comput. Sci. 18, 5 (2024), 185603

    2024

  3. [11]

    McAuley, and Caiming Xiong

    Yongjun Chen, Zhiwei Liu, Jia Li, Julian J. McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. In WWW ’22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022 . ACM, 2172–2182

    2022

  4. [12]

    Yankai Chen, Quoc-Tuan Truong, Xin Shen, Jin Li, and Irwin King. 2024. Shopping Trajectory Representation Learning with Pre-training for E-commerce Customer Understanding and Recommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Minin...

    2024

  5. [13]

    Sihao Ding, Fuli Feng, Xiangnan He, Yong Liao, Jun Shi, and Yongdong Zhang

  6. [14]

    Jing Du, Zesheng Ye, Bin Guo, Zhiwen Yu, and Lina Yao. 2024. Identifiability of Cross-Domain Recommendation via Causal Subspace Disentanglement. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, Wash...

    2024

  7. [15]

    Yingpeng Du, Ziyan Wang, Zhu Sun, Yining Ma, Hongzhi Liu, and Jie Zhang

  8. [16]

    Chen Gao, Yu Zheng, Wenjie Wang, Fuli Feng, Xiangnan He, and Yong Li. 2024. Causal Inference in Recommender Systems: A Survey and Future Directions. ACM Trans. Inf. Syst. 42, 4 (2024), 88:1–88:32

    2024

  9. [17]

    Shen Gao, Jiabao Fang, Quan Tu, Zhitao Yao, Zhumin Chen, Pengjie Ren, and Zhaochun Ren. 2024. Generative News Recommendation. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 . ACM, 3444–3453

    2024

  10. [18]

    Xiang Gao, Meera Sitharam, and Adrian E Roitberg. 2017. Bounds on the Jensen gap, and implications for mean-concentrated distributions. arXiv preprint arXiv:1712.05267 (2017)

    2017 arXiv

  11. [19]

    In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024

    Disentangled Multi-interest Representation Learning for Sequential Rec- ommendation. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelona, Spain, August 25-29, 2024 . ACM, 677–688

    2024

  12. [20]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). In RecSys ’22: Sixteenth ACM Conference on Recommender Systems, Seattle, W A, USA, Septem...

    2022

  13. [21]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In Proceedings of the 43rd International ACM SIGIR con- ference on research and development in Information Retriev...

    2020

  14. [22]

    Jin Huang, Harrie Oosterhuis, Masoud Mansoury, Herke van Hoof, and Maarten de Rijke. 2024. Going Beyond Popularity and Positivity Bias: Correcting for Mul- tifactorial Bias in Recommender Systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and De...

    2024

  15. [23]

    Yingqiang Ge, Shuya Zhao, Honglu Zhou, Changhua Pei, Fei Sun, Wenwu Ou, and Yongfeng Zhang. 2020. Understanding Echo Chambers in E-commerce Rec- ommender Systems. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval...

    2020

  16. [24]

    Wei Jiang, Xinyi Gao, Guandong Xu, Tong Chen, and Hongzhi Yin. 2024. Chal- lenging Low Homophily in Social Recommendation. InProceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 . ACM, 3476–3484

    2024

  17. [25]

    Jiazheng Jing, Yinan Zhang, Xin Zhou, and Zhiqi Shen. 2023. Capturing Pop- ularity Trends: A Simplistic Non-Personalized Approach for Enhanced Item Recommendation. In Proceedings of the 32nd ACM International Conference on In- formation and Knowledge Management, CIKM 2023, Bir...

    2023

  18. [26]

    Thorsten Joachims, Adith Swaminathan, and Tobias Schnabel. 2018. Unbiased Learning-to-Rank with Biased Feedback. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI 2018, July 13-19, 2018, Stockholm, Sweden. ijcai.org, 5284–5288

    2018

  19. [27]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques. ACM Trans. Inf. Syst. 20, 4 (2002), 422–446

    2002

  20. [28]

    Dawen Liang, Laurent Charlin, and David M Blei. 2016. Causal inference for recommendation. In Causation: Foundation to Application, Workshop at UAI. AUAI, Vol. 6. 108

    2016

  21. [29]

    Siyi Lin, Chongming Gao, Jiawei Chen, Sheng Zhou, Binbin Hu, Yan Feng, Chun Chen, and Can Wang. 2025. How Do Recommendation Models Amplify Popularity Bias? An Analysis from the Spectral Perspective. In Proceedings of the Eighteenth ACM International Conference on Web Search an...

    2025

  22. [30]

    Fangyuan Luo and Jun Wu. 2023. User-Dependent Learning to Debias for Rec- ommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023. ACM, 2491–2495

    2023

  23. [31]

    McAuley, and Jure Leskovec

    Himabindu Lakkaraju, Julian J. McAuley, and Jure Leskovec. 2013. What’s in a Name? Understanding the Interplay between Titles, Content, and Communities in Social Media. In Proceedings of the Seventh International Conference on Weblogs and Social Media, ICWSM 2013, Cambridge, M...

    2013

  24. [32]

    Judith Möller, Damian Trilling, Natali Helberger, and Bram van Es. 2020. Do not blame it on the algorithm: an empirical assessment of multiple recommender sys- tems and their impact on content diversity. InDigital media, political polarization and challenges to democracy . Rou...

    2020

  25. [33]

    Ashutosh Nayak, Mayur Garg, and Rajasekhara Reddy Duvvuru Muni. 2023. News Popularity Beyond the Click-Through-Rate for Personalized Recommendations. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023,...

    2023

  26. [34]

    Nguyen, Pik-Mai Hui, F

    Tien T. Nguyen, Pik-Mai Hui, F. Maxwell Harper, Loren G. Terveen, and Joseph A. Konstan. 2014. Exploring the filter bubble: the effect of using recommender systems on content diversity. In 23rd International World Wide Web Conference, WWW ’14, Seoul, Republic of Korea, April 7...

    2014

  27. [35]

    Huishi Luo, Fuzhen Zhuang, Ruobing Xie, Hengshu Zhu, Deqing Wang, Zhulin An, and Yongjun Xu. 2024. A survey on causal inference for recommendation. The Innovation (2024)

    2024

  28. [36]

    Junwei Pan, Wei Xue, Ximei Wang, Haibin Yu, Xun Liu, Shijie Quan, Xueming Qiu, Dapeng Liu, Lei Xiao, and Jie Jiang. 2024. Ads Recommendation in a Collapsed and Entangled World. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, B...

    2024

  29. [37]

    J Pearl. 2009. Causality. Cambridge university press

    2009

  30. [38]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme

  31. [39]

    Wentao Ning, Reynold Cheng, Xiao Yan, Ben Kao, Nan Huo, Nur Al Hasan Haldar, and Bo Tang. 2024. Debiasing Recommendation with Personal Popularity. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024. ACM, 3400–3409

    2024

  32. [40]

    Robert Sanders. 1987. The Pareto principle: its use and abuse. Journal of Services Marketing 1, 2 (1987), 37–40

    1987

  33. [41]

    Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, and Thorsten Joachims. 2016. Recommendations as Treatments: Debiasing Learning and Evaluation. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19...

    2016

  34. [42]

    Yu Shang, Chen Gao, Jiansheng Chen, Depeng Jin, and Yong Li. 2024. Improving Item-side Fairness of Multimodal Recommendation via Modality Debiasing. In Proceedings of the ACM Web Conference 2024 . 4697–4705

    2024

  35. [43]

    Tianhao Shi, Yang Zhang, Jizhi Zhang, Fuli Feng, and Xiangnan He. 2024. Fair Recommendations with Limited Sensitive Attributes: A Distributionally Robust Optimization Approach. InProceedings of the 47th International ACM SIGIR Confer- ence on Research and Development in Inform...

    2024

  36. [44]

    Yuta Saito, Suguru Yaginuma, Yuta Nishino, Hayato Sakata, and Kazuhide Nakata

  37. [45]

    In WSDM ’20: The Thirteenth ACM International Conference on Web Search and Data Mining, Houston, TX, USA, February 3-7, 2020

    Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback. In WSDM ’20: The Thirteenth ACM International Conference on Web Search and Data Mining, Houston, TX, USA, February 3-7, 2020 . ACM, 501–509. Taming Recommendation Bias with Causal Intervention on Evolv...

    2020

  38. [46]

    Shisong Tang, Qing Li, Dingmin Wang, Ci Gao, Wentao Xiao, Dan Zhao, Yong Jiang, Qian Ma, and Aoyang Zhang. 2023. Counterfactual Video Recommendation for Duration Debiasing. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2023, Long ...

    2023

  39. [47]

    Hangyu Wang, Jianghao Lin, Bo Chen, Yang Yang, Ruiming Tang, Weinan Zhang, and Yong Yu. 2025. Towards efficient and effective unlearning of large language models for recommendation. Frontiers Comput. Sci. 19, 3 (2025), 193327

    2025

  40. [48]

    Jizhe Wang, Pipei Huang, Huan Zhao, Zhibo Zhang, Binqiang Zhao, and Dik Lun Lee. 2018. Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD 2018, Lon...

    2018

  41. [49]

    Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, and Tat-Seng Chua. 2021. Deconfounded Recommendation for Alleviating Bias Amplification. In KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18, 2021 . ACM, 1717–1725

    2021

  42. [50]

    Weiping Song, Zhiping Xiao, Yifan Wang, Laurent Charlin, Ming Zhang, and Jian Tang. 2019. Session-Based Social Recommendation via Dynamic Graph Attention Networks. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM 2019, Melbourne, V...

    2019

  43. [51]

    Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang

  44. [52]

    Tianxin Wei, Fuli Feng, Jiawei Chen, Ziwei Wu, Jinfeng Yi, and Xiangnan He

  45. [53]

    Wei Wei, Jiabin Tang, Lianghao Xia, Yangqin Jiang, and Chao Huang. 2024. PromptMM: Multi-Modal Knowledge Distillation for Recommendation with Prompt-Tuning. In Proceedings of the ACM on Web Conference 2024, WWW 2024, Singapore, May 13-17, 2024 . ACM, 3217–3228

    2024

  46. [54]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen

  47. [55]

    Xiaolong Xu, Hongsheng Dong, Lianyong Qi, Xuyun Zhang, Haolong Xiang, Xiaoyu Xia, Yanwei Xu, and Wanchun Dou. 2024. CMCLRec: Cross-modal Con- trastive Learning for User Cold-start Sequential Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Resea...

    2024

  48. [56]

    Chen Yang, Sunhao Dai, Yupeng Hou, Wayne Xin Zhao, Jun Xu, Yang Song, and Hengshu Zhu. 2024. Revisiting Reciprocal Recommender Systems: Metrics, Formulation, and Method. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, KDD 2024, Barcelon...

    2024

  49. [57]

    Xiangmeng Wang, Qian Li, Dianer Yu, Peng Cui, Zhichao Wang, and Guan- dong Xu. 2023. Causal Disentanglement for Semantic-Aware Intent Learning in Recommendation. IEEE Trans. Knowl. Data Eng. 35, 10 (2023), 9836–9849

    2023

  50. [58]

    Yixin Wang, Dawen Liang, Laurent Charlin, and David M Blei. 2018. The decon- founded recommender: A causal inference approach to recommendation. arXiv preprint arXiv:1808.06581 (2018)

    2018 arXiv

  51. [59]

    An Zhang, Jingnan Zheng, Xiang Wang, Yancheng Yuan, and Tat-Seng Chua. 2023. Invariant Collaborative Filtering to Popularity Distribution Shift. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023, Ying Ding, Jie Tang, Juan F. S...

    2023

  52. [60]

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. 2021. Causal Intervention for Leveraging Popularity Bias in Recommendation. In SIGIR ’21: The 44th International ACM SIGIR Conference on Research and Development in Information Re...

    2021

  53. [61]

    Minghao Zhao, Le Wu, Yile Liang, Lei Chen, Jian Zhang, Qilin Deng, Kai Wang, Xudong Shen, Tangjie Lv, and Runze Wu. 2022. Investigating Accuracy-Novelty Performance for Graph-based Collaborative Filtering. In SIGIR ’22: The 45th International ACM SIGIR Conference on Research a...

    2022

  54. [62]

    Zihao Zhao, Jiawei Chen, Sheng Zhou, Xiangnan He, Xuezhi Cao, Fuzheng Zhang, and Wei Wu. 2023. Popularity Bias is not Always Evil: Disentangling Benign and Harmful Bias for Recommendation. IEEE Trans. Knowl. Data Eng. 35, 10 (2023), 9920–9931

    2023

  55. [63]

    World Wide Web (WWW) 27, 5 (2024), 60

    A survey on large language models for recommendation. World Wide Web (WWW) 27, 5 (2024), 60

    2024

  56. [64]

    Yu Zheng, Chen Gao, Xiang Li, Xiangnan He, Yong Li, and Depeng Jin. 2021. Disentangling User Interest and Conformity for Recommendation with Causal Embedding. In WWW ’21: The Web Conference 2021, Virtual Event / Ljubljana, Slovenia, April 19-23, 2021. ACM / IW3C2, 2980–2991

    2021

  57. [65]

    Kun Zhou, Hui Wang, Wayne Xin Zhao, Yutao Zhu, Sirui Wang, Fuzheng Zhang, Zhongyuan Wang, and Ji-Rong Wen. 2020. S3-Rec: Self-Supervised Learning for Sequential Recommendation with Mutual Information Maximization. In CIKM ’20: The 29th ACM International Conference on Informati...

    2020

  58. [66]

    Yuhao Yang, Chao Huang, Lianghao Xia, Chunzhen Huang, Da Luo, and Kangyi Lin. 2023. Debiased Contrastive Learning for Sequential Recommendation. In Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023 . ACM, 1063–1073

    2023

  59. [67]

    An Zhang, Wenchang Ma, Jingnan Zheng, Xiang Wang, and Tat-Seng Chua. 2024. Robust Collaborative Filtering to Popularity Distribution Shift. ACM Transactions on Information Systems 42, 3 (Jan. 2024), 1–25

    2024

  60. [72]

    Yu Zheng, Chen Gao, Jianxin Chang, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. 2022. Disentangling Long and Short-Term Interests for Recommendation. In WWW ’22: The ACM Web Conference 2022, Virtual Event, Lyon, France, April 25 - 29, 2022. ACM, 2256–2267

    2022

  61. [75]

    Xinyuan Zhu, Yang Zhang, Fuli Feng, Xun Yang, Dingxian Wang, and Xiangnan He. 2024. Mitigating Hidden Confounding Effects for Causal Recommendation. IEEE Trans. Knowl. Data Eng. 36, 9 (2024), 4794–4805

    2024

  62. [76]

    Ziwei Zhu, Yun He, Xing Zhao, Yin Zhang, Jianling Wang, and James Caverlee

  63. [77]

    In WSDM ’21, The Fourteenth ACM International Conference on Web Search and Data Mining, Virtual Event, Israel, March 8-12, 2021

    Popularity-Opportunity Bias in Collaborative Filtering. In WSDM ’21, The Fourteenth ACM International Conference on Web Search and Data Mining, Virtual Event, Israel, March 8-12, 2021 . ACM, 85–93

    2021

  64. [2009]

    In UAI 2009, Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, Montreal, QC, Canada, June 18-21, 2009

    BPR: Bayesian Personalized Ranking from Implicit Feedback. In UAI 2009, Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence, Montreal, QC, Canada, June 18-21, 2009 . AUAI Press, 452–461

    2009

  65. [2019]

    In Proceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019

    BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. In Proceedings of the 28th ACM International Con- ference on Information and Knowledge Management, CIKM 2019, Beijing, China, November 3-7, 2019. ACM, 1441–1450

    2019

  66. [2020]

    In RecSys 2020: Fourteenth ACM Conference on Recommender Systems, Virtual Event, Brazil, September 22-26, 2020

    The Connection Between Popularity Bias, Calibration, and Fairness in Recommendation. In RecSys 2020: Fourteenth ACM Conference on Recommender Systems, Virtual Event, Brazil, September 22-26, 2020 . ACM, 726–731

    2020

  67. [2021]

    In KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18,

    Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System. In KDD ’21: The 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, Singapore, August 14-18,

  68. [2023]

    ACM Trans

    Bias and Debias in Recommender System: A Survey and Future Directions. ACM Trans. Inf. Syst. 41, 3 (2023), 67:1–67:39

    2023

  69. [2024]

    IEEE Trans

    Causal Incremental Graph Convolution for Recommender System Retrain- ing. IEEE Trans. Neural Networks Learn. Syst. 35, 4 (2024), 4718–4728

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.