Pith. sign in

REVIEW 3 major objections 8 minor 62 references

Addressing Correlated Latent Exogenous Variables in Debiased Recommender Systems

T0 review · 3 major / 8 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 This paper claims that selection bias in recommender data can be corrected without additional randomized trials by modeling the unobserved noise behind observation and rating as a correlated pair rather than as independent variables.

desk verdict A genuinely new take on an old bias problem, with correct likelihood math, but the unbiasedness claim is only proven for the pure likelihood, not for the algorithm that is actually run. read the letter →

arxiv 2506.07517 v1 pith:4B6JVN26 submitted 2025-06-09 cs.LG cs.IR

Shuqiang Zhang , Yuchao Zhang , Jinkun Chen , Haochen Sui This is my paper
classification cs.LGcs.IR
keywords debiasedrecommendationselectionbiaslatentexogenousvariableslikelihoodmaximizationMonteCarloestimationmissingnotatrandombivariatenormalerrorsunmeasuredconfounding
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

Recommender systems learn from user interactions that are missing not at random, since users mostly interact with items they already prefer. Standard debiasing estimators—error imputation, inverse propensity scoring, and doubly robust methods—assume the unobserved noise driving who responds and what rating they give is independent. This paper argues that those two noise sources can be correlated in practice, and proposes a likelihood-based learner that estimates the correlation together with the prediction model. Under a bivariate normal assumption on the latent noise, the method recovers unbiased predictions without needing randomized control data. Experiments on synthetic data and on Coat, Yahoo! R3, and KuaiRec show stable performance across correlation values and gains over existing debiasing baselines.

What carries the argument

The central object is the bivariate normal pair of latent exogenous variables $(U_O, U_R)$ with correlation $\rho$, added to the observation score $z_{u,i}=g_o(x_{u,i};\theta_o)+U_O$ and the rating score $y_{u,i}=g_r(x_{u,i};\theta_r)+U_R$. The machinery is the resulting likelihood: continuous ratings give a closed form involving $\Phi$, and binary ratings give $f(z>0,y>0)=\mathbb{E}_\epsilon[\Phi((g_r+\rho\epsilon)/\sqrt{1-\rho^2}) I\{\epsilon>-g_o\}]$, estimated with Monte Carlo samples. A symmetry property—swapping $g_o$ and $g_r$ in the integrand yields the companion estimate—allows alternating optimization of the propensity model and the prediction model while sharing $\rho$, bypassing the non-differentiable indicator function.

What would settle it

Generate synthetic data with the same threshold structure but draw the two noise terms from a $t$-distribution, or from normals whose variance depends on the features, or with a correlation that varies across user groups, then fit the proposed likelihood and check whether the estimated $\rho$ and test-set predictive error track the truth.

Watch

Extended reading notes

Core claim

The paper's central claim is that the standard independence assumption on exogenous variables in recommender-system structural causal models can be relaxed if the data-generation process is modeled directly. It treats the observation score and the rating score as additive functions of observed features plus two latent noise terms $U_O$ and $U_R$ that follow a bivariate normal distribution with mean zero, unit variance, and correlation $\rho$. The likelihood of the observed missing-not-at-random data is then expressible through the normal CDF $\Phi$ for continuous ratings, and for binary ratings it becomes a one-dimensional Gaussian expectation that can be estimated by Monte Carlo sampling. The authors exploit a symmetry between the propensity and prediction models to alternate their optimization while sharing $\rho$, avoiding the non-differentiable indicator in the binary likelihood. The paper claims this yields an unbiased prediction model without additional RCT data, and demonstrates the claim on synthetic data with known $\rho$ and on three real-world recommendation datasets.

Load-bearing premise

The whole method rests on Assumption 1: the two unobserved noise variables are bivariate normal with mean zero, variance one, and a single correlation, added to the observation and rating scores; if the real noise is non-normal, heteroskedastic, or group-dependent, the likelihood is misspecified and the claimed unbiasedness fails.

Editorial extensions

If this is right

  • Debiasing no longer requires a separate randomized control dataset when the normality assumption holds: the likelihood itself estimates the hidden coupling $\rho$ and corrects it.
  • The estimated $\rho$ gives practitioners a diagnostic for how strongly unobserved observation noise and preference noise are coupled in their data.
  • The likelihood term can be blended with any existing debiasing loss through a single weight $\alpha$, so the method composes with IPS, DR, and related estimators instead of replacing them.
  • The same derivation covers continuous outcomes such as watch time or purchase amount exactly and binary outcomes such as clicks approximately by Monte Carlo.

Reading between the lines

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

  • Not tested in the paper: generating data with the same thresholds but non-normal, heteroskedastic, or group-varying noise would directly probe whether the bivariate-normal assumption is load-bearing or merely convenient.
  • The symmetry trick used here suggests a wider design pattern for two-sided missingness problems, such as survey nonresponse or clinical registry linkage, where two latent scores plausibly share a correlation parameter.
  • Because correlated $U_O$ and $U_R$ can be re-expressed as a common hidden cause, the method's practical gain may come as much from parametric efficiency as from the shift in causal interpretation.
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

3 major / 8 minor

Summary. The manuscript proposes a generative model for user-item interaction data in which the latent exogenous variables driving the observation indicator O and the rating R are allowed to be correlated, in contrast to the standard independence assumption in prior debiasing methods. The authors derive the likelihood for continuous feedback and a Monte Carlo approximation for binary feedback, then introduce an alternating optimization scheme exploiting an exchange symmetry of the joint probability. The final objective combines the (negative) log-likelihood with an arbitrary debiasing loss via a weight α. Experiments on a semi-synthetic MovieLens setup and on Coat, Yahoo! R3, and KuaiRec show gains over a wide range of baselines. The paper claims that the method 'can achieve unbiasedness without additional RCT data' (Section 4.1).

Significance. The likelihood derivation is clear and the Monte Carlo estimator in Eq. (12) is unbiased for the target joint probability, and the appendices provide a detailed computation. If the consistency claim held for the actual algorithm, this would be a useful contribution: it relaxes a key assumption of the IPS/DR framework and offers a principled joint learning scheme for the propensity and outcome models. The code is released, and the experimental coverage is broad. However, the gap between the likelihood theory and the implemented objective means the headline theoretical property is not yet established for the estimator that produces the reported numbers.

major comments (3)
  1. [Section 4.1, Eq. (15), Algorithm 1] The paper claims that the proposed method 'can achieve unbiasedness without additional RCT data' (Section 4.1). This claim is only substantiated for the pure likelihood objective (10). The deployed Algorithm 1 minimizes the weighted combination L_E = α(−L_B_MLE) + (1−α)L_Debias (Eq. 15) with α ∈ (0,1) tuned per dataset, as shown in Figure 4. For a generic debiasing loss, the expected gradient at the true parameters is generally nonzero unless the propensity or imputation model is exactly correct, so the minimizer of L_E is shifted from the true g_r. No theorem or argument is given that the mixture preserves consistency under Assumption 1; this is a load-bearing gap because the real-world results in Table 2 are obtained with this mixture. The authors should either prove a condition under which the mixture is still consistent, or restrict the unbiasedness claim to the pure likelihood variant and report its results separately.
  2. [Section 4.4, Eqs. (12)–(14)] The alternating decomposition of the binary likelihood is not equivalent to maximizing L_B_MLE. Both MC_r (Eq. 12) and MC_o (Eq. 13) estimate the same joint probability f(z>0,y>0), but the two sub-objectives L_R_MLE and L_D_MLE replace that term by different stochastic estimates, and the algorithm alternates between them. This is a coordinate-descent heuristic for which no convergence or consistency proof is supplied. In addition, because log MC is a biased estimate of the log-likelihood for any finite L (Jensen's inequality), the 'likelihood' objective is not an unbiased estimator of the true log-likelihood; this bias is not discussed. The paper should provide an asymptotic argument (e.g., L →∞ with a suitable step-size schedule, or concentration of the MC error), or explicitly characterize the approximation error and its effect on the estimator.
  3. [Section 5 and Table 2] The empirical validation does not separate the effect of the likelihood term from the debiasing loss. In the semi-synthetic experiments of Section 5, the data are generated exactly from the bivariate normal additive-noise model of Assumption 1, so they do not test sensitivity to misspecification. The real-world experiments in Table 2 compare against baselines but do not report the tuned α values, nor do they include the α=1 (pure likelihood) variant, which is the estimator for which the unbiasedness claim is derived. To substantiate the paper's central claim, the authors should report results for α=1, report the chosen α values, and include a misspecification experiment (e.g., non-normal or heteroskedastic latent noises, or ρ varying across user-item pairs).
minor comments (8)
  1. [Throughout] The paper uses 'unbiasedness' where the maximum likelihood estimator is at best consistent, not unbiased in finite samples. Please adjust the terminology (e.g., Section 4.1 and the abstract) to avoid overclaiming.
  2. [Section 4.2, Eq. (9) and Appendix A] The squared term in exp{−(y−g_r)^2/2} is typeset incorrectly (e.g., 'exp{−(y_u,i−g_r(x_u,i;θ_r)2) 2}' in Eq. (9) and similar in Eq. (38)). Please fix the parenthesization.
  3. [Section 4.3] The notation R under the sums denotes the set of observed positive ratings, but R is also used for the full rating matrix in Section 3. Please use a distinct symbol and define it.
  4. [Section 4.5 and Algorithm 1] Eq. (15) defines L_E using L_B_MLE, but Algorithm 1 actually optimizes α(−L_R_MLE)+(1−α)L_Debias for the prediction model and −L_D_MLE for the propensity model. Please clarify how the alternating steps relate to the combined objective in Eq. (15).
  5. [Section 6.2.2 and Figure 3] The text says the Monte Carlo sample size is tuned in [100,500] for Coat, but Figure 3 shows the Coat x-axis running from 300 to 700. Please reconcile the reported range with the figure.
  6. [Table 2] The baseline labeled MRDR-DL is referred to as MRDR in the text; please align the naming.
  7. [Table 3] The running time for 'Ours' is reported without specifying which variant (Ours-Naive or Ours-DR) is measured. Please clarify.
  8. [Abstract and Section 1] The phrase 'we release this assumption' appears in the abstract and in Section 1; the intended wording is 'relax'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the likelihood derivation is self-contained from explicit normality assumptions; self-citations are not load-bearing.

full rationale

The derivation chain is self-contained. Assumption 1 explicitly postulates (U_O, U_R) bivariate normal with mean 0, variance 1, covariance ρ, and defines z = g_o(x;θ_o) + U_O, y = g_r(x;θ_r) + U_R. Equation (7) is the resulting bivariate normal density, Eq. (8) is the marginal probit probability for non-observation, Eq. (9) is obtained by direct integration of z over (0,∞), and Eq. (10) is the negative log-likelihood to minimize. The binary-preference MC estimator in Eqs. (11)-(13) follows from the same density by another explicit integration and the stated symmetry of the two arguments; the alternation in Eq. (14) is a coordinate-wise optimization of that likelihood. Nowhere is the prediction target g_r (or the estimated ρ) used to define the likelihood or the exogenous distribution, so there is no self-definitional or fitted-input-called-prediction step. The self-citations present (e.g., [19], [26], [53] involving one of the authors) appear only as general contextual references for time series, information overload, and bias-variance trade-off; they carry none of the load of the central derivation, and no uniqueness theorem from the authors' prior work is invoked to force the model choice. The paper's normality assumption is stated as an assumption, not smuggled through citation. The potential gap noted by the skeptic—that Algorithm 1 minimizes α(−L_MLE) + (1−α)L_Debias and unbiasedness is only argued for the pure likelihood objective—is a correctness/robustness issue about the mixed estimator, not a circularity: the debiasing loss does not define the likelihood or the target. Consequently, the central claim does not reduce to its inputs by construction.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the bivariate normality and additive-noise assumptions, which are strong and unvalidated. The symmetry property is standard mathematics. No new entities are postulated.

free parameters (3)
  • ρ (correlation between U_O and U_R) = estimated during training
    Central parameter of the bivariate normal latent variable model, estimated by maximum likelihood in both synthetic and real experiments.
  • α (trade-off weight) = tuned per dataset
    Balances the likelihood loss and the debiasing loss in Eq. (15); Section 6.6 sensitivity analysis shows α substantially affects performance.
  • Monte Carlo sample size L = 100-500 (Coat), 10-60 (Yahoo, KuaiRec)
    Tuned per dataset; affects accuracy and variance of the likelihood estimate, as shown in Section 6.5.
assumptions (3)
  • ad hoc to paper Assumption 1: U_O and U_R follow a bivariate normal distribution with mean 0, variance 1, and covariance ρ.
    Introduced for tractability; central to the likelihood derivation and not verified on real data.
  • domain assumption Additive noise structure: z = g_o(x) + U_O and y = g_r(x) + U_R.
    Assumes the latent variables enter additively; a standard assumption in sample selection models but not grounded in the recommender setting.
  • standard math Symmetry of the bivariate normal CDF, used to split the optimization into L_R and L_D.
    Mathematically true, but the finite-sample Monte Carlo estimates are not exactly symmetric, which is a source of approximation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Addressing Correlated Latent Exogenous Variables in Debiased Recommender Systems." pith.science (2026). https://pith.science/paper/4B6JVN26

@misc{pith2026250607517,
  author       = {Pith},
  title        = {Pith review of: Addressing Correlated Latent Exogenous Variables in Debiased Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4B6JVN26}},
  note         = {Machine review of arXiv:2506.07517}
}
read the original abstract

Recommendation systems (RS) aim to provide personalized content, but they face a challenge in unbiased learning due to selection bias, where users only interact with items they prefer. This bias leads to a distorted representation of user preferences, which hinders the accuracy and fairness of recommendations. To address the issue, various methods such as error imputation based, inverse propensity scoring, and doubly robust techniques have been developed. Despite the progress, from the structural causal model perspective, previous debiasing methods in RS assume the independence of the exogenous variables. In this paper, we release this assumption and propose a learning algorithm based on likelihood maximization to learn a prediction model. We first discuss the correlation and difference between unmeasured confounding and our scenario, then we propose a unified method that effectively handles latent exogenous variables. Specifically, our method models the data generation process with latent exogenous variables under mild normality assumptions. We then develop a Monte Carlo algorithm to numerically estimate the likelihood function. Extensive experiments on synthetic datasets and three real-world datasets demonstrate the effectiveness of our proposed method. The code is at https://github.com/WallaceSUI/kdd25-background-variable.

Figures

Figures reproduced from arXiv: 2506.07517 by the authors.

Figure 1
Figure 1. Comparison between traditional unmeasured con [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Estimation of 𝜌 with different ground truth. to increase as 𝜌 becomes smaller. During these conditions, the IPS and DR methods outperform our proposed method. This is because that when 𝜌 is small, the bias introduced by the correlated latent ex￾ogenous variables is also small that can not dominate the total bias. This finding supports the incorporation of conventional debiasing methods to enhance predictive performa… view at source ↗
Figure 3
Figure 3. Performance compared to baseline with varying Monte Carlo sample sizes. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Model performance for different value of trade-off parameter [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 57 canonical work pages

  1. [1]

    Jiangxia Cao, Jiawei Sheng, Xin Cong, Tingwen Liu, and Bin Wang. 2022. Cross- Domain Recommendation to Cold-Start Users via Variational Information Bottle- neck. InInternational Conference on Data Engineering

    work page 2022

  2. [2]

    Jiawei Chen, Hande Dong, Yang Qiu, Xiangnan He, Xin Xin, Liang Chen, Guli Lin, and Keping Yang. 2021. Autodebias: Learning to Debias for Recommendation. In International ACM SIGIR Conference on Research and Development in Information Retrieval

    work page 2021

  3. [3]

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

  4. [4]

    Zhichao Chen, Hao Wang, Zhihuan Song, and Zhiqiang Ge. 2024. Improving Data-Driven Inferential Sensor Modeling by Industrial Knowledge: A Bayesian Perspective.IEEE Transactions on Systems, Man, and Cybernetics: Systems(2024)

    work page 2024

  5. [5]

    Quanyu Dai, Haoxuan Li, Peng Wu, Zhenhua Dong, Xiao-Hua Zhou, Rui Zhang, Xiuqiang He, Rui Zhang, and Jie Sun. 2022. A Generalized Doubly Robust Learning Framework for Debiasing Post-Click Conversion Rate Prediction. In ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    work page 2022

  6. [6]

    Sihao Ding, Peng Wu, Fuli Feng, Yitong Wang, Xiangnan He, Yong Liao, and Yong- dong Zhang. 2022. Addressing Unmeasured Confounder for Recommendation with Sensitivity Analysis. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    work page 2022

  7. [7]

    Chongming Gao, Shijun Li, Wenqiang Lei, Jiawei Chen, Biao Li, Peng Jiang, Xiangnan He, Jiaxin Mao, and Tat-Seng Chua. 2022. KuaiRec: A Fully-Observed Dataset and Insights for Evaluating Recommender Systems. InACM International Conference on Information & Knowledge Management

    work page 2022

  8. [8]

    Siyuan Guo, Lixin Zou, Yiding Liu, Wenwen Ye, Suqi Cheng, Shuaiqiang Wang, Hechang Chen, Dawei Yin, and Yi Chang. 2021. Enhanced Doubly Robust Learn- ing for Debiasing Post-Click Conversion Rate Estimation. InInternational ACM SIGIR Conference on Research and Development in Information Retrieval

    work page 2021

Show all 62 references
  1. [9]

    Mingming Ha, Taoxuewen, Wenfang Lin, Qiongxu Ma, Wujiang Xu, and Linxun Chen. 2024. Fine-Grained Dynamic Framework for Bias-Variance Joint Opti- mization on Data Missing Not at Random. InAdvances in Neural Information Processing Systems

    2024

  2. [10]

    Joseph Y Halpern and Evan Piermont. 2024. Subjective Causality.arXiv preprint arXiv:2401.10937(2024)

    2024 arXiv

  3. [11]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural Collaborative Filtering. InInternational Conference on World Wide Web

    2017

  4. [12]

    Shanshan Huang, Haoxuan Li, Chunyuan Zheng, Mingyuan Ge, Wei Gao, Lei Wang, and Li Liu. 2025. Text-Driven Fashion Image Editing with Compositional Concept Learning and Counterfactual Abduction. InIEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  5. [13]

    Shanshan Huang, Haoxuan Li, Chunyuan Zheng, Lei Wang, Guorui Liao, Zhili Gong, Huayi Yang, and Li Liu. 2025. Visual Representation Learning through Causal Intervention for Controllable Image Editing. InIEEE/CVF Conference on Computer Vision and Pattern Recognition

    2025

  6. [14]

    Aapo Hyvärinen and Petteri Pajunen. 1999. Nonlinear Independent Component Analysis: Existence and Uniqueness Results.Neural Networks12, 3 (1999), 429– 439

    1999

  7. [15]

    Adrián Javaloy, Pablo Sánchez-Martín, and Isabel Valera. 2024. Causal Normaliz- ing Flows: From Theory to Practice. InAdvances in Neural Information Processing Systems

    2024

  8. [16]

    Shuoran Jiang, Qingcai Chen, Yang Xiang, Youcheng Pan, Xiangping Wu, and Yukang Lin. 2024. Confounder Balancing in Adversarial Domain Adaptation for Pre-Trained Large Models Fine-Tuning.Neural Networks173 (2024), 106173

    2024

  9. [17]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Tech- niques for Recommender Systems.Computer42, 8 (2009), 30–37

    2009

  10. [18]

    Wonbin Kweon and Hwanjo Yu. 2024. Doubly Calibrated Estimator for Recom- mendation on Data Missing Not At Random. InInternational World Wide Web Conference

    2024

  11. [19]

    Songning Lai, Ninghui Feng, Haochen Sui, Ze Ma, Hao Wang, Zichen Song, Hang Zhao, and Yutao Yue. 2024. FTS: A Framework to Find a Faithful TimeSieve. arXiv preprint arXiv:2405.19647(2024)

    2024 arXiv

  12. [20]

    Haoxuan Li, Yan Lyu, Chunyuan Zheng, and Peng Wu. 2023. TDR-CL: Tar- geted Doubly Robust Collaborative Learning for Debiased Recommendations. In International Conference on Learning Representations

    2023

  13. [21]

    Haoxuan Li, Kunhan Wu, Chunyuan Zheng, Yanghao Xiao, Hao Wang, Zhi Geng, Fuli Feng, Xiangnan He, and Peng Wu. 2023. Removing Hidden Confounding in Recommendation: A Unified Multi-Task Learning Approach. InAdvances in Neural Information Processing Systems

    2023

  14. [22]

    Haoxuan Li, Yanghao Xiao, Chunyuan Zheng, and Peng Wu. 2023. Balancing Unobserved Confounding with a Few Unbiased Ratings in Debiased Recommen- dations. InInternational World Wide Web Conference

    2023

  15. [23]

    Haoxuan Li, Yanghao Xiao, Chunyuan Zheng, Peng Wu, and Peng Cui. 2023. Propensity Matters: Measuring and Enhancing Balancing for Recommendation. InInternational Conference on Machine Learning

    2023

  16. [24]

    Haoxuan Li, Chunyuan Zheng, and Peng Wu. 2023. StableDR: Stabilized Dou- bly Robust Learning for Recommendation on Data Missing Not at Random. In International Conference on Learning Representations

    2023

  17. [25]

    Haoxuan Li, Chunyuan Zheng, Yanghao Xiao, Peng Wu, Zhi Geng, Xu Chen, and Peng Cui. 2024. Debiased Collaborative Filtering with Kernel-based Causal Balancing. InInternational Conference on Learning Representations

    2024

  18. [26]

    Meng Li and Haochen Sui. 2025. Causal Recommendation via Machine Unlearning with a Few Unbiased Data. InAAAI Workshop on Artificial Intelligence with Causal Techniques

    2025

  19. [27]

    Dugang Liu, Pengxiang Cheng, Zinan Lin, Jinwei Luo, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming. 2022. KDCRec: Knowledge Distillation for Coun- terfactual Recommendation via Uniform Data.IEEE Transactions on Knowledge and Data Engineering(2022)

    2022

  20. [28]

    Dugang Liu, Pengxiang Cheng, Hong Zhu, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming. 2021. Mitigating Confounding Bias in Recommendation via Information Bottleneck. InACM Conference on Recommender Systems

    2021

  21. [29]

    Weiming Liu, Chaochao Chen, Xinting Liao, Mengling Hu, Jiajie Su, Yanchao Tan, and Fan Wang. 2024. User Distribution Mapping Modelling with Collaborative Filtering for Cross Domain Recommendation. InInternational World Wide Web Conference

    2024

  22. [30]

    Jinwei Luo, Dugang Liu, Weike Pan, and Zhong Ming. 2021. Unbiased Recom- mendation Model Based on Improved Propensity Score Estimation.Journal of Computer Applications41, 12 (2021), 3508

    2021

  23. [31]

    Jing Ma, Mengting Wan, Longqi Yang, Jundong Li, Brent Hecht, and Jaime Teevan

  24. [32]

    Benjamin M Marlin and Richard S Zemel. 2009. Collaborative Prediction and Ranking with Non-random Missing Data. InACM Conference on Recommender Systems

    2009

  25. [33]

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. 2021. Normalizing Flows for Probabilistic Model- ing and Inference.Journal of Machine Learning Research22, 57 (2021), 1–64

    2021

  26. [34]

    2009.Causality

    Judea Pearl. 2009.Causality. Cambridge university press

    2009

  27. [35]

    Yuta Saito. 2020. Asymmetric Tri-training for Debiasing Missing-Not-At-Random Explicit Feedback. InInternational ACM SIGIR Conference on Research and Devel- opment in Information Retrieval

    2020

  28. [36]

    Yuta Saito. 2020. Doubly Robust Estimator for Ranking Metrics with Post-Click Conversions. InACM Conference on Recommender Systems

    2020

  29. [37]

    Yuta Saito and Masahiro Nomura. 2022. Towards Resolving Propensity Contra- diction in Offline Recommender Learning. InInternational Joint Conference on Artificial Intelligence

    2022

  30. [38]

    Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, and Thorsten Joachims. 2016. Recommendations as Treatments: Debiasing Learning and Evaluation. InInternational Conference on Machine Learning

    2016

  31. [39]

    Harald Steck. 2010. Training and Testing of Recommender Systems on Data Missing Not at Random. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2010

  32. [40]

    Fan Wang, Chaochao Chen, Weiming Liu, Tianhao Fan, Xinting Liao, Yanchao Tan, Lianyong Qi, and Xiaolin Zheng. 2024. CE-RCFR: Robust Counterfactual Regression for Consensus-Enabled Treatment Effect Estimation. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2024

  33. [41]

    Fan Wang, Lianyong Qi, Weiming Liu, Bowen Yu, Jintao Chen, and Yanwei Xu

  34. [42]

    Hao Wang. 2024. Improving Neural Network Generalization on Data-Limited Regression with Doubly-Robust Boosting. InAAAI Conference on Artificial Intel- ligence

    2024

  35. [43]

    Hao Wang, Tai-Wei Chang, Tianqiao Liu, Jianmin Huang, Zhichao Chen, Chao Yu, Ruopeng Li, and Wei Chu. 2022. ESCM2: Entire Space Counterfactual Multi-task Model for Post-Click Conversion Rate Estimation. InInternational ACM SIGIR Conference on Research and Development in Inform...

    2022

  36. [44]

    Hao Wang, Zhichao Chen, Zhaoran Liu, Xu Chen, Haoxuan Li, and Zhouchen Lin. 2025. Proximity Matters: Local Proximity Enhanced Balancing for Treatment Effect Estimation. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    2025

  37. [45]

    Hao Wang, Zhichao Chen, Zhaoran Liu, Haozhe Li, Degui Yang, Xinggao Liu, and Haoxuan Li. 2024. Entire Space Counterfactual Learning for Reliable Content Recommendations.IEEE Transactions on Information Forensics and Security20 (2024), 1755–1764

    2024

  38. [46]

    Hao Wang, Jiajun Fan, Zhichao Chen, Haoxuan Li, Weiming Liu, Tianqiao Liu, Quanyu Dai, Yichao Wang, Zhenhua Dong, and Ruiming Tang. 2023. Optimal Transport for Treatment Effect Estimation. InAdvances in Neural Information Processing Systems

    2023

  39. [47]

    Haotian Wang, Haoxuan Li, Hao Zou, Haoang Chi, Long Lan, Wanrong Huang, and Wenjing Yang. 2025. Effective and Efficient Time-Varying Counterfactual Prediction with State-Space Models. InInternational Conference on Learning Representations. KDD ’25, August 3–7, 2025, Toronto, O...

    2025

  40. [48]

    Wenjie Wang, Yang Zhang, Haoxuan Li, Peng Wu, Fuli Feng, and Xiangnan He

  41. [49]

    Xiaojie Wang, Rui Zhang, Yu Sun, and Jianzhong Qi. 2019. Doubly Robust Joint Learning for Recommendation on Data Missing Not at Random. InInternational Conference on Machine Learning

    2019

  42. [50]

    Xiaojie Wang, Rui Zhang, Yu Sun, and Jianzhong Qi. 2021. Combating selection biases in recommender systems with a few unbiased ratings. InACM International Conference on Web Search and Data Mining

    2021

  43. [51]

    Zifeng Wang, Xi Chen, Rui Wen, Shao-Lun Huang, Ercan Kuruoglu, and Yefeng Zheng. 2020. Information Theoretic Counterfactual Learning from Missing-Not- at-Random Feedback. InAdvances in Neural Information Processing Systems

    2020

  44. [52]

    InInternational ACM SIGIR Conference on Research and Development in Information Retrieval

    Causal Recommendation: Progresses and Future Directions. InInternational ACM SIGIR Conference on Research and Development in Information Retrieval

  45. [53]

    Yuntian Wu, Yuntian Yang, Jiabao Sean Xiao, Chuan Zhou, Haochen Sui, and Haoxuan Li. 2024. Invariant Spatiotemporal Representation Learning for Cross- patient Seizure Classification. InNeurIPS Workshop on NeuroAI

    2024

  46. [54]

    Quanhan Xi and Benjamin Bloem-Reddy. 2023. Indeterminacy in Generative Models: Characterization and Strong Identifiability. InInternational Conference on Artificial Intelligence and Statistics

    2023

  47. [55]

    Yanghao Xiao, Haoxuan Li, Yongqiang Tang, and Wensheng Zhang. 2024. Ad- dressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation. InAdvances in Neural Information Processing Systems

    2024

  48. [56]

    Peng Wu, Haoxuan Li, Yuhao Deng, Wenjie Hu, Quanyu Dai, Zhenhua Dong, Jie Sun, Rui Zhang, and Xiao-Hua Zhou. 2022. On the Opportunity of Causal Learning in Recommendation Systems: Foundation, Estimation, Prediction and Challenges.arXiv preprint arXiv:2201.06716(2022)

    2022 arXiv

  49. [57]

    Chunyuan Zheng, Hang Pan, Yang Zhang, and Haoxuan Li. 2025. Adaptive Structure Learning with Partial Parameter Sharing for Post-Click Conversion Rate Prediction. InInternational ACM SIGIR Conference on Research and Development in Information Retrieval

    2025

  50. [58]

    Φ(𝑔𝑜(𝑥;𝜃 𝑜)+𝜌𝑝√︁ 1−𝜌 2 )𝐼{𝑝>−𝑔 𝑟(𝑥;𝜃 𝑟)} # (56) =𝑓(𝑧 𝑢,𝑖 >0,𝑦 𝑢,𝑖 >0)(57) =E𝜙(𝑝)

    Chuan Zhou, Yaxuan Li, Chunyuan Zheng, Haiteng Zhang, Min Zhang, Haoxuan Li, and Mingming Gong. 2025. A Two-Stage Pretraining-Finetuning Framework for Treatment Effect Estimation with Unmeasured Confounding. InACM SIGKDD Conference on Knowledge Discovery and Data Mining. Addre...

    2025

  51. [60]

    Wenhao Zhang, Wentian Bao, Xiao-Yang Liu, Keping Yang, Quan Lin, Hong Wen, and Ramin Ramezani. 2020. Large-scale Causal Approaches to Debiasing Post-click Conversion Rate Estimation with Multi-task Learning. InInternational World Wide Web Conference

    2020

  52. [2022]

    InACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Learning Causal Effects on Hypergraphs. InACM SIGKDD Conference on Knowledge Discovery and Data Mining

  53. [2023]

    ACM Transactions on Information Systems41, 3 (2023), 1–39

    Bias and Debias in Recommender System: A Survey and Future Directions. ACM Transactions on Information Systems41, 3 (2023), 1–39

    2023

  54. [2025]

    Inter- and Intra- Similarity Preserved Counterfactual Incentive Effect Estimation for Recommendation Systems.ACM Transactions on Information Systems(2025)

    2025

Pith tools

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