Pith. sign in

REVIEW 5 major objections 5 minor 38 references

Conditional, instance-level alignment between biased and unbiased towers can debias recommendations without losing factual accuracy—and is deployable in streaming industrial systems.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-07-31 23:04 UTC pith:CY3NQ3YV

load-bearing objection Clever gated alignment with real online A/B evidence, but the MAR assumption and missing offline error bars need tighter support. the 5 major comments →

arxiv 2607.24092 v1 pith:CY3NQ3YV submitted 2026-07-27 cs.IR

ConAlign: Conditional Alignment Framework for Balancing Biased and Unbiased Recommendation

Jingcheng Zhang , Yihan Wang , Qi Song , Liyin Hong This is my paper
classification cs.IR
keywords recommender systemsdebiasingfilter bubblerandom exposureknowledge distillationdual-tower architecturestreaming trainingonline A/B testing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to solve a practical tension: recommender systems trained on logged interactions are biased and create filter bubbles, but switching to debiased training often hurts the metrics that matter in the real serving environment. ConAlign's answer is a dual-tower design where a tiny unbiased tower, trained on a small stream of random exposures, produces the final ranking, and a large biased tower, trained on normal logs, is used only as a guide. The guide is gated: the unbiased tower is pulled toward the biased tower's latent representation only when the biased tower's prediction on that sample has lower loss. The paper argues that this selective transfer preserves factual accuracy while still capturing true preferences, and it backs the claim with offline results on three public datasets and an online A/B test.

Core claim

ConAlign's core claim is that a discrete gate—align only when the biased tower's loss is lower—makes it possible to train a debiased tower that beats prior debiasing methods on unbiased test sets without losing ground in biased settings. The alignment happens on penultimate hidden representations rather than final scores, giving the unbiased tower richer collaborative knowledge while stop-gradients keep the biased tower's training unaffected. On Yahoo! R3, ConAlign reports unbiased UAUC 0.6841 versus InterD's 0.6787; on Coat and KuaiRand-Pure it also leads the unbiased columns, while keeping biased performance competitive with the best baseline. In live deployment on a large short-video plat

What carries the argument

The central object is the conditional alignment loss with the gating indicator I_cond = [ℓ_b < ℓ_d]: for each training instance, if the biased tower's prediction loss is lower than the unbiased tower's, the unbiased tower's penultimate hidden state is pulled toward the biased tower's (with stop-gradient on the biased side). This turns what would be a universal knowledge-distillation into a selective intervention, letting the unbiased tower learn freely except where the biased tower clearly knows better. The gating is evaluated on mixed mini-batches during streaming training, and the unbiased tower itself is served at inference.

Load-bearing premise

Everything rests on the unbiased stream being genuinely unbiased: the paper equates uniform random insertion with missing-at-random preference evidence, but attention and context effects on a randomly placed video can depress clicks, and the stream is only a tiny fraction of traffic.

What would settle it

A concrete test: on a held-out set of items, compare ConAlign's top-ranked items with those chosen by a much larger randomized study (e.g., every user rates 100 uniformly sampled videos). If ConAlign's ranking diverges from the random-evidence ranking in a way that tracks the biased tower's preferences, the MAR assumption or the gating mechanism is failing. A simpler check is to run ConAlign with the gate always off on a large unbiased dataset; if performance does not drop, the gate is not actually doing the work.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Platforms can deploy debiasing with a negligible fraction of random traffic; the intervention rate is kept small enough not to disturb the feed, yet the unbiased tower benefits from the biased tower's full-scale knowledge.
  • Because alignment is at the representation level and gated per instance, the framework avoids the expensive bi-level meta-learning and full Cartesian-product distillation of prior methods, making streaming training feasible.
  • The balanced performance suggests debiasing need not trade off short-term engagement; the reported A/B gains imply reduced filter bubbles can coincide with stable or improved DAU/LT7.
  • The streaming formulation lets the unbiased tower track shifting user interests and item popularity in real time, which static unbiased datasets cannot do.
  • Serving the unbiased tower's predictions directly, rather than blending predictions, is what allows the system to actually show diversified content, a design choice other methods could adopt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The per-instance loss comparison is a crude confidence signal; a learned or temperature-scaled gate might transfer knowledge more smoothly and avoid the sharp threshold where performance drops when alignment weight is too high (which the paper's hyperparameter analysis shows).
  • Because the unbiased tower's input includes the biased tower's penultimate representation, bias is never fully excised; if random-exposure clicks are themselves contextually depressed (attention, insertion position), the method may re-inject a shrunken version of the very bias it aims to remove. A large purely-random dataset would reveal whether the gate is actually recovering true preferences.
  • The online gains, though statistically significant, are tiny in absolute terms (+0.069% DAU); the paper's argument that these are practically significant relies on retention being notoriously hard to move, but an independent replication on another platform is needed before generalizing the effect size.
  • One natural extension is to apply the same conditional alignment to other biases separately (position, popularity, selection), turning the binary gate into a multi-dimensional switch.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes ConAlign, a dual-tower recommendation framework that trains a biased tower on observational feedback and an unbiased tower on a small stream of uniformly randomized exposure data, with a conditional gating mechanism that selectively aligns the unbiased tower's representations to the biased tower when the biased tower has lower loss. The claim is that this preserves factual (biased-environment) accuracy while improving unbiased preference estimation. Offline experiments on Coat, Yahoo! R3, and KuaiRand-Pure report consistent gains in UAUC/NDCG@5 over baselines including InterD, and a 10% traffic live A/B on Kuaishou reports small but statistically significant gains in DAU, LT7, and diversity metrics.

Significance. If the claims hold, ConAlign is a practically relevant industrial debiasing framework: it is lightweight (training time 122.7s vs. 10592.1s for InterD on Yahoo! R3), supports streaming training, and the online A/B provides an external validation that is not dependent on the paper's own hyperparameter choices. The paper also releases code and does not rely on self-citations. However, the central debiasing claim rests on the assertion in §3.3.1 that uniformly sampled random exposure yields MAR data, and the offline evidence has no error bars or significance tests; the key ablation gain is +0.0007 UAUC. These issues are load-bearing and need to be addressed.

major comments (5)
  1. [§3.3.1] The MAR claim is load-bearing: the paper asserts that because one video is uniformly sampled from the candidate pool and inserted at a uniformly random position, the resulting interaction signals satisfy MAR. Uniform exposure removes selection bias in what items are shown, but the observed outcome (click, valid_play, watch) is still a context-dependent behavioral response. A video injected mid-feed, outside the user's current session context, is likely to receive depressed engagement relative to the same item served organically. The paper gives no diagnostic validating that random-exposure engagement rates are unbiased preference measurements. It also deliberately keeps the intervention probability small, so the unbiased tower trains on a negligible traffic fraction. This threatens the validity of the unbiased tower and the online diversity/lift results. Please provide (i) a calibration
  2. [Table 1 and Table 2] All offline results are single runs with no error bars, standard deviations, or significance tests. This is particularly problematic for the central claim of consistent improvement. In Table 2, the alignment-loss ablation (ours-noAlign vs. ConAlign) on unbiased UAUC is 0.6834 vs. 0.6841, a gain of +0.0007; this is within typical run-to-run noise for deep models. Similarly, the margin over InterD on unbiased Yahoo UAUC (0.6841 vs. 0.6787) is 0.0054, and on NDCG@5 (0.7761 vs. 0.7713) is 0.0048 — meaningful only if accompanied by variance estimates. Please report multiple seeds and paired significance tests (e.g., bootstrap or paired t-test) for at least the main comparisons and the ablation table.
  3. [§3.2.2, Eq. (7)] The gating condition I_cond = [ℓ_b < ℓ_d] is underspecified. The mixed mini-batch consists of both biased and unbiased interactions; it is not clear which label is used to compute ℓ_b on an unbiased instance and ℓ_d on a biased instance. If ℓ_b is computed using the biased label while ℓ_d uses the unbiased label, the comparison is apples-to-oranges. If both are computed on the same (unbiased) label, then the gate measures the biased tower's performance on unbiased data, which has different semantics. Moreover, the paper never reports the fraction of instances in which the gate is active, how it evolves during training, or its stability across batches. This gate is the key novelty; without this analysis the 'conditional' mechanism is not empirically characterized. Please clarify the loss computation and report gate activation statistics.
  4. [§4.4, Table 3] The online A/B gains are small (DAU +0.069%, LT7 +0.029%) but reported with CIs. The diversity metrics VIN and CC are computed from user feedback; if the random-intervention video injections themselves contribute to these metrics (e.g., a randomly shown video leading to a new valid interest), the favorable diversity shift could partly reflect the intervention mechanism rather than the model's debiasing. The paper does not state whether the random intervention probability and the tagging of injected items are identical across the A/B groups, or whether VIN/CC are computed excluding randomly injected items. Please clarify the online experimental design and report diversity metrics excluding intervention traffic as a robustness check.
  5. [§4.2] The paper claims 'ConAlign consistently outperforms all baselines in unbiased recommendation performance across all datasets.' This is overstated given the lack of variance estimates (see above). In particular, on Coat the unbiased NDCG gain over InterD is 0.7066 vs. 0.6939, while on KuaiRand-Pure the unbiased NDCG gain is 0.5656 vs. 0.5544; but without error bars, 'consistently' is not established. Please temper the claim or provide statistical support.
minor comments (5)
  1. [§4.1.2, Baselines] The baseline name is listed as 'CauseE' in the main text but 'CausE' in Table 1 and in the related-work section. Please unify the spelling.
  2. [§4.1.3, Implementation Details] The hyperparameter search space is written as '{1, 0.1, 0.01, 0.001, 0.00001, 0.000001}' without indicating which hyperparameters these values apply to (learning rate and weight decay presumably). Please make the grid explicit per hyperparameter.
  3. [Table 1 caption] The caption says 'The last column reports the training latency on Yahoo! R3' but the column is positioned after the KuaiRand-Pure columns. It would be clearer to label it explicitly as 'Time (s) — Yahoo! R3' and place it as a separate table or clearly separated column.
  4. [§4.3.1, Ablation Study] In the text, the variant name 'ours-biasTower' is used in Table 2 but described as 'Only the biased tower of ConAlign is retained.' This is clear, but please ensure the naming is consistent in Figure 4/5 captions and text.
  5. [§3.3.2, Streaming Training] The training details are sparse: no batch sizes, learning rates for the online streaming setting, or how model updates are scheduled (e.g., per-batch vs. per-window). Adding these would strengthen reproducibility, especially since the paper claims industrial streaming deployment.

Circularity Check

0 steps flagged

No circularity: ConAlign's central claims are empirical and are validated on external benchmarks and live A/B; the MAR assertion is an assumption, not a circular derivation.

full rationale

ConAlign does not derive its predictions from fitted constants or from a self-citation chain. The gating mechanism in Eq. (7) is operationally defined by comparing biased-tower and unbiased-tower losses, and the alignment loss in Eq. (6) is a regularizer; the final recommendation output is the unbiased tower's score, and the paper's claimed gains are measured on held-out unbiased test sets (Coat, Yahoo! R3, KuaiRand-Pure) and a live Kuaishou A/B test. No fitted parameter is renamed as a prediction, and no author self-citation supplies a load-bearing uniqueness or existence argument. The MAR claim in §3.3.1 is an assumption about the random-intervention stream, not a step that reduces to its own inputs; a violation of MAR would weaken the validity of the unbiased stream, but that is an empirical concern, not circularity. The internal offline metrics are omitted for business sensitivity, which is a reporting limitation, not a circular step. Overall, the derivation chain is self-contained: the method is an architecture plus training objective, and its evaluation is external.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

ConAlign's central claim rests on three domain assumptions: (1) the random-insertion stream is missing-at-random, making feedback on random items a valid estimate of P(R|D); (2) the biased tower's penultimate representation, concatenated into the unbiased input (Eq. 3), carries transferable semantic knowledge rather than irremovable policy bias; (3) representation-space closeness (Eq. 6) transfers capacity without transferring biased probabilities. None is derived or stress-tested. Free parameters include the alignment weight λ (Eq. 8), the undisclosed random-intervention probability (§3.3.1), and the per-batch biased/unbiased sample counts (§3.3.2). No invented entities beyond architectural components; the "latent bias representation" is a named hidden-layer vector, not an independent entity with a falsifiable handle.

free parameters (3)
  • alignment weight λ = not reported (swept, Fig. 4-5)
    Scales the conditional alignment loss (Eq. 8) and controls the biased/unbiased trade-off on which the paper's central balance claim depends; chosen per dataset by search.
  • random-intervention probability = not disclosed ("a fixed probability", §3.3.1)
    Determines the size and quality of the unbiased MAR stream, which is the paper's core data input; the value is withheld.
  • per-batch mixed-stream composition = not disclosed ("a fixed number of examples from each stream", §3.3.2)
    Controls the biased/unbiased supervision ratio in streaming training; no reported value or sensitivity analysis.
axioms (5)
  • domain assumption The random-insertion protocol produces MAR interaction data, so feedback on randomly inserted items estimates P(R|D).
    §3.3.1 asserts the MAR condition follows from uniform item selection and uniform position insertion; never verified against stated preferences or attention confounds. Load-bearing because the unbiased tower's entire value is the unbiasedness of this stream.
  • domain assumption The biased tower's penultimate representation h_b^(L-1) is a useful input for unbiased preference estimation.
    Eq. (3) concatenates the biased last hidden layer into the unbiased tower's input; assumes policy bias is not irreversibly baked into this representation. Not tested, despite being the main route by which biased signals enter the "unbiased" tower.
  • domain assumption Penultimate-layer MSE alignment (Eq. 6) transfers knowledge without transferring biased probabilities.
    Standard distillation-style assumption, asserted in §3.2.2 with no support; the Table 2 ablation shows alignment mainly rescues biased-environment performance, not unbiased accuracy.
  • domain assumption Binary thresholding of Coat and Yahoo! R3 ratings (>3 = positive), inherited from InterD [5].
    Data preprocessing adopted from prior work (§4.1.1); not a paper contribution but affects all reported numbers.
  • standard math Supervised losses (Eq. 2, Eq. 5) and stop-gradient training are valid learning setups.
    Backpropagation and BCE-style losses are standard; no issues.

pith-pipeline@v1.3.0-alltime-deepseek · 12712 in / 19024 out tokens · 162214 ms · 2026-07-31T23:04:14.657458+00:00 · methodology

0 comments
read the original abstract

Industry recommender systems trained on observational data suffer from various biases that create filter bubbles, causing user interests to collapse into narrow categories and severely degrading long-term engagement. While utilizing unbiased uniform data for debiasing has shown promise, existing methods remain impractical for industrial deployment due to limitations such as neglect of factual (biased) recommendation performance and the substantial computational overhead. To overcome these limitations, we propose ConAlign (Conditional Alignment Framework), a conditional debiasing approach for industrial deployment. The key innovation of ConAlign lies in a discrete gating-based conditional alignment mechanism that selectively transfers knowledge from the biased tower to the unbiased tower. Following a selective intervention paradigm rather than universal correction, it seamlessly balances factual accuracy and unbiased preference estimation while supporting real-time streaming adaptation. To the best of our knowledge, ConAlign is the first streaming debiasing recommendation framework successfully deployed in a large-scale industrial recommendation system that utilizes a small fraction of unbiased random traffic for debiasing. Extensive offline experiments on three real-world datasets rigorously validate the effectiveness of our proposed framework. Furthermore, large-scale online A/B testing on Kuaishou demonstrates significant improvements in long-term user engagement and interest diversity, with negligible latency overhead.

Figures

Figures reproduced from arXiv: 2607.24092 by Jingcheng Zhang, Liyin Hong, Qi Song, Yihan Wang.

Figure 1
Figure 1. Figure 1: The filter bubble phenomenon in biased recommender systems. Nodes represent latent user interests; [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The correlation between Valid Interest Number [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the ConAlign framework. ConAlign consists of a biased tower trained on biased interaction data and [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UAUC performance of ConAlign under different [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: NDCG@5 performance of ConAlign under different [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 4 canonical work pages · 4 internal anchors

  1. [1]

    Ting Bai, Weijie Chen, Cheng Yang, and Chuan Shi. 2025. Invariant Debiasing Learning for Recommendation via Biased Imputation. arXiv:2412.20036 [cs] doi:10.48550/arXiv.2412.20036

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.20036 2025

  2. [3]

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

  3. [4]

    Andrew Collins, Dominika Tkaczyk, Akiko Aizawa, and Joeran Beel

  4. [5]

    Sihao Ding, Fuli Feng, Xiangnan He, Jinqiu Jin, Wenjie Wang, Yong Liao, and Yongdong Zhang. 2022. Interpolative Distillation for Unifying Biased and De- biased Recommendation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, Madrid Spain, 40–49. doi:10.1145/3477495.3532002

    arXiv 2022

  5. [6]

    Chongming Gao, Shijun Li, Yuan Zhang, Jiawei Chen, Biao Li, Wenqiang Lei, Peng Jiang, and Xiangnan He. 2022. KuaiRand: An Unbiased Sequential Rec- ommendation Dataset with Randomly Exposed Videos. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 3953–3957. arXiv:2208.08696 [cs] doi:10.1145/3511808.3557624

    Pith/arXiv arXiv 2022

  6. [7]

    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. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Informa- tion Retrieval. ACM, Virtual Event Canada, 275–28...

    arXiv 2021

  7. [8]

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the Knowledge in a Neural Network. arXiv:1503.02531 [stat] doi:10.48550/arXiv.1503.02531

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.02531 2015

  8. [9]

    Ray Jiang, Silvia Chiappa, Tor Lattimore, András György, and Pushmeet Kohli

  9. [10]

    Marius Kaminskas and Derek Bridge. 2017. Diversity, Serendipity, Novelty, and Coverage: A Survey and Empirical Analysis of Beyond-Accuracy Objectives in Recommender Systems.ACM Transactions on Interactive Intelligent Systems7, 1 (March 2017), 1–42. doi:10.1145/2926720

    doi:10.1145/2926720 2017

  10. [11]

    Yehuda Koren, Robert Bell, and Chris Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems.Computer42, 8 (Aug. 2009), 30–37. doi:10.1109/MC.2009.263

    doi:10.1109/mc.2009.263 2009

  11. [12]

    Adit Krishnan, Ashish Sharma, Aravind Sankar, and Hari Sundaram. 2018. An Adversarial Approach to Improve Long-Tail Performance in Neural Collaborative Filtering. InProceedings of the 27th ACM International Conference on Information and Knowledge Management. ACM, Torino Italy, 1491–1494. doi:10.1145/3269206. 3269264

    doi:10.1145/3269206 2018

  12. [13]

    Haoxuan Li, Yan Lyu, Chunyuan Zheng, and Peng Wu. 2023. TDR-CL: Tar- geted Doubly Robust Collaborative Learning for Debiased Recommendations. arXiv:2203.10258 [cs] doi:10.48550/arXiv.2203.10258

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2203.10258 2023

  13. [14]

    Haoxuan Li, Chunyuan Zheng, Shuyi Wang, Kunhan Wu, Eric Wang, Peng Wu, Zhi Geng, Xu Chen, and Xiao-Hua Zhou. 2024. Relaxing the Accurate Imputation Assumption in Doubly Robust Learning for Debiased Collaborative Filtering. In Forty-First International Conference on Machine Learning

    2024

  14. [15]

    Meng Li and Haochen Sui. 2025. Causal Recommendation via Machine Unlearning with a Few Unbiased Data. InAAAI 2025 Workshop on Artificial Intelligence with Causal Techniques, Vol. 2

    2025

  15. [16]

    Dawen Liang, Laurent Charlin, James McInerney, and David M. Blei. 2016. Mod- eling User Exposure in Recommendation. InProceedings of the 25th International Conference on World Wide Web. International World Wide Web Conferences Steer- ing Committee, Montréal Québec Canada, 951–961. doi:10.1145/2872427.2883090

    arXiv 2016

  16. [17]

    Dugang Liu, Pengxiang Cheng, Zhenhua Dong, Xiuqiang He, Weike Pan, and Zhong Ming. 2020. A General Knowledge Distillation Framework for Counterfac- tual Recommendation via Uniform Data. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, Virtual Event China, 831–840. doi:10.1145/3397271.3401083

    arXiv 2020

  17. [18]

    Haochen Liu, Da Tang, Ji Yang, Xiangyu Zhao, Hui Liu, Jiliang Tang, and Youlong Cheng. 2022. Rating Distribution Calibration for Selection Bias Mitigation in Recommendations. InProceedings of the ACM Web Conference 2022 (WWW ’22). Association for Computing Machinery, New York, NY, USA, 2048–2057. doi:10. 1145/3485447.3512078

    arXiv 2022

  18. [19]

    Masoud Mansoury, Himan Abdollahpouri, Mykola Pechenizkiy, Bamshad Mobasher, and Robin Burke. 2020. Feedback Loop and Bias Amplification in Recommender Systems. arXiv:2007.13019 [cs] doi:10.48550/arXiv.2007.13019

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2007.13019 2020

  19. [20]

    Zemel, Sam Roweis, and Malcolm Slaney

    Benjamin Marlin, Richard S. Zemel, Sam Roweis, and Malcolm Slaney. 2012. Col- laborative Filtering and the Missing at Random Assumption. arXiv:1206.5267 [cs] doi:10.48550/arXiv.1206.5267

    Pith/arXiv arXiv doi:10.48550/arxiv.1206.5267 2012

  20. [21]

    Marlin and Richard S

    Benjamin M. Marlin and Richard S. Zemel. 2009. Collaborative Prediction and Ranking with Non-Random Missing Data. InProceedings of the Third ACM Conference on Recommender Systems. ACM, New York New York USA, 5–12. doi:10.1145/1639714.1639717

    arXiv 2009

  21. [22]

    Fred Rowland. 2011. The Filter Bubble: What the Internet Is Hiding from You. portal: Libraries and the Academy11, 4 (2011), 1009–1011

    2011

  22. [23]

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

  23. [24]

    Tobias Schnabel, Adith Swaminathan, Ashudeep Singh, Navin Chandak, and Thorsten Joachims. 2016. Recommendations as Treatments: Debiasing Learning and Evaluation. InInternational Conference on Machine Learning. PMLR, 1670– 1679

    2016

  24. [25]

    Harald Steck. 2013. Evaluation of Recommendations: Rating-Prediction and Ranking. InProceedings of the 7th ACM Conference on Recommender Systems. ACM, Hong Kong China, 213–220. doi:10.1145/2507157.2507160

    arXiv 2013

  25. [26]

    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. PMLR, 6638–6647

    2019

  26. [27]

    Xiaojie Wang, Rui Zhang, Yu Sun, and Jianzhong Qi. 2021. Combating Selection Biases in Recommender Systems with a Few Unbiased Ratings. InProceedings of the 14th ACM International Conference on Web Search and Data Mining (WSDM ’21). Association for Computing Machinery, New York, NY, USA, 427–435. doi:10. 1145/3437963.3441799

    arXiv 2021

  27. [28]

    Yu, and Peng Cui

    Zimu Wang, Yue He, Jiashuo Liu, Wenchao Zou, Philip S. Yu, and Peng Cui. 2022. Invariant Preference Learning for General Debiasing in Recommendation. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’22). Association for Computing Machinery, New York, NY, USA, 1969–1978. doi:10.1145/3534678.3539439

    arXiv 2022

  28. [29]

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

  29. [30]

    Xinyang Yi, Ji Yang, Lichan Hong, Derek Zhiyuan Cheng, Lukasz Heldt, Aditee Kumthekar, Zhe Zhao, Li Wei, and Ed Chi. 2019. Sampling-Bias-Corrected Neural Modeling for Large Corpus Item Recommendations. InProceedings of the 13th ACM Conference on Recommender Systems (RecSys ’19). Association for Computing Machinery, New York, NY, USA, 269–277. doi:10.1145/...

    arXiv 2019

  30. [31]

    An Zhang, Wenchang Ma, Pengbo Wei, Leheng Sheng, and Xiang Wang. 2024. General Debiasing for Graph-based Collaborative Filtering via Adversarial Graph Dropout. https://arxiv.org/abs/2402.13769v1

    Pith/arXiv arXiv 2024

  31. [32]

    Qing Zhang, Xiaoying Zhang, Yang Liu, Hongning Wang, Min Gao, Jiheng Zhang, and Ruocheng Guo. 2023. Debiasing Recommendation by Learning Identifiable Latent Confounders. https://arxiv.org/abs/2302.05052v2

    Pith/arXiv arXiv 2023

  32. [33]

    Kuiyu Zhu, Tao Qin, Pinghui Wang, and Xin Wang. [n. d.]. Adversarial Propensity Weighting for Debiasing in Collaborative Filtering. ([n. d.])

  33. [34]

    McNee, Joseph A

    Cai-Nicolas Ziegler, Sean M. McNee, Joseph A. Konstan, and Georg Lausen. 2005. Improving Recommendation Lists through Topic Diversification. InProceedings of the 14th International Conference on World Wide Web - WWW ’05. ACM Press, Chiba, Japan, 22. doi:10.1145/1060745.1060754

    arXiv 2005

  34. [2018]

    https://arxiv.org/abs/1802.06565v1

    A Study of Position Bias in Digital Library Recommender Systems. https://arxiv.org/abs/1802.06565v1

    Pith/arXiv arXiv

  35. [2019]

    InProceed- ings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society

    Degenerate Feedback Loops in Recommender Systems. InProceed- ings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. 383–390. arXiv:1902.10730 [stat] doi:10.1145/3306618.3314288

    Pith/arXiv arXiv 2019

  36. [2020]

    Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback

    Unbiased Recommender Learning from Missing-Not-At-Random Implicit Feedback. arXiv:1909.03601 [stat] doi:10.48550/arXiv.1909.03601

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1909.03601 1909

  37. [2021]

    InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining

    Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. ACM, Virtual Event Singapore, 1791–1800. doi:10.1145/3447548.3467289

    arXiv

  38. [2023]

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

    Bias and Debias in Recommender System: A Survey and Future Directions. ACM Transactions on Information Systems41, 3 (July 2023), 1–39. doi:10.1145/ 3564284

    2023