Pith. sign in

REVIEW 3 major objections 5 minor 56 references

A plug-in causal framework debiases multi-behavior recommenders by intervening on user habits and item engagement patterns, then fusing auxiliaries with MoE gates and bias-aware contrastive alignment.

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 · grok-4.5

2026-07-13 18:25 UTC pith:2DRK2ICF

load-bearing objection Solid plug-in for multi-behavior rec that delivers consistent lifts; the causal story is mostly packaging around frequency proxies and the real work is MoE fusion + bias-aware CL. the 3 major comments →

arxiv 2603.25126 v2 pith:2DRK2ICF submitted 2026-03-26 cs.IR cs.AI

MCLMR: A Model-Agnostic Causal Learning Framework for Multi-Behavior Recommendation

Ranxu Zhang , Junjie Meng , Ying Sun , Ziqi Xu , Bing Yin , Hao Li , Yanyong Zhang , Chao Wang This is my paper
classification cs.IR cs.AI
keywords Recommender SystemsCausal InterventionMulti-behaviorMixture-of-ExpertsContrastive LearningDebiasingModel-agnostic
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.

Multi-behavior recommenders use views, carts, and purchases to fight sparsity, but they routinely confuse true preference with user habits and uneven item engagement patterns. This paper claims those confounders can be blocked by a causal graph plus do-intervention, then residual auxiliary signals can be fused dynamically and aligned across behaviors. The resulting model-agnostic module, MCLMR, drops into existing multi-behavior architectures and lifts hit-rate and NDCG on three public e-commerce logs. A sympathetic reader cares because the same plug-in both removes popular multi-behavior distortions and turns noisy auxiliaries into useful transfer without redesigning the backbone.

Core claim

MCLMR shows that intervening on a multi-behavior causal graph (user and item bias nodes as confounders of mediating interactions and target outcomes) yields unbiased preference scores; when those scores are further refined by a Mixture-of-Experts adaptive aggregator and a bias-aware dual-view contrastive loss, diverse multi-behavior backbones improve significantly on Tmall, Jdata and Taobao.

What carries the argument

Causal Preference Estimation via backdoor adjustment on the multi-behavior graph (parameterized by relative frequency proxies for user/item biases), paired with MoE semantic gating plus Jaccard structural gating for auxiliary fusion and InfoNCE with per-user/item bias-dependent temperature for cross-behavior alignment.

Load-bearing premise

Relative frequencies of each behavior type for a user or item are good enough proxies for the hidden confounders that the backdoor formula actually removes the bias.

What would settle it

Replace the relative-frequency bias proxies with random noise or with absolute counts and re-run the full suite on Tmall/Jdata/Taobao; if the reported HR@10/NDCG@10 gains over the same backbones vanish or reverse, the central causal claim fails.

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

If this is right

  • Any multi-behavior GNN or factorization backbone can be wrapped by MCLMR and should inherit measurable ranking gains without architectural redesign.
  • Recommendations become less skewed toward high-auxiliary, low-decision items once the dual confounders are intervened on.
  • Less-active users receive larger relative lifts than active users, partially countering the Matthew effect of sparse histories.
  • Inference latency remains identical to the backbone because all causal modules act only at training time.
  • Bias-aware temperature can be reused in other multi-view contrastive recommenders that face heterogeneous reliability across views.

Where Pith is reading between the lines

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

  • The same relative-frequency proxy plus MoE gate pattern could be ported to multi-modal or multi-intent recommenders where different signal sources act as latent confounders.
  • If the proxy assumption is the soft spot, future work could learn the confounder distributions jointly rather than fixing them to conversion rates, testing whether residual bias still leaks into the interventional scores.
  • The dual user/item contrastive views suggest a natural extension to session-level or sequential multi-behavior graphs where temporal order supplies an extra conditioning variable.

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

3 major / 5 minor

Summary. The paper proposes MCLMR, a model-agnostic plug-in for multi-behavior recommendation that (i) builds a causal graph treating user multi-behavior habits and item multi-behavior distributions as confounders, (ii) parameterizes a backdoor adjustment with relative-frequency bias proxies, (iii) fuses auxiliary behaviors via a dual-path Adaptive Aggregation module (Jaccard structural gate + MoE semantic gate), and (iv) aligns cross-behavior embeddings with a bias-aware InfoNCE temperature. At inference it ranks by the inner product of target-behavior embeddings. Experiments on Tmall, Jdata, and Taobao show consistent HR@K/NDCG@K gains when MCLMR is attached to LightGCN, CRGCN, BCIPM, and HEC-GCN, with ablations, user/item group analyses, complexity bounds, and runtime measurements.

Significance. If the empirical gains hold under broader scrutiny, MCLMR is a practically useful contribution: a drop-in causal-style module that improves both simple and strong multi-behavior backbones, with public code and three standard datasets. The dual-path aggregation and bias-aware contrastive temperature are concrete, reusable design choices. The work is less decisive as a causal identification result: the unbiasedness claim rests on relative-frequency proxies for latent confounders, which the authors themselves flag as an open approximation. The main value for the community is therefore the engineering framework and the broad, reproducible empirical package rather than a fully identified causal estimator.

major comments (3)
  1. [§4.1.2, Eqs. (3)–(5)] §4.1.2, Eq. (3) and the parameterization in Eqs. (4)–(5): the central claim of “unbiased preference estimation” via backdoor adjustment depends on relative frequencies b_{u,k} and b_{i,k} being adequate proxies for latent confounders B_u and B_i. The manuscript acknowledges that optimal proxies remain open and follows prior convention, but supplies no diagnostic (sensitivity to alternative propensity estimators, correlation with external habit measures, or placebo interventions). Without this, the do-calculus derivation supports a modeling story more than a verified identification result; the reported lifts could be driven largely by MoE fusion and bias-aware CL. Please either (a) add proxy-sensitivity / alternative-propensity experiments, or (b) systematically soften “unbiased” language to “bias-aware / approximately deconfounded” and state the identification assumptions explicitly in t
  2. [§4.2, Eq. (6); §4.4, Eq. (13)] §4.2 vs §4.4: training optimizes a debiased score S_final,K_t that multiplies base scores by adaptive contributions (Eq. 6), yet inference ranks solely by the plain inner product of target-behavior embeddings (Eq. 13). This training–inference mismatch weakens the link between the interventional objective and the scores that are actually evaluated. Clarify whether the causal adjustment is intended only as a training regularizer, and report an ablation that ranks with S_final (or a frozen debiased scorer) at test time so readers can see how much of the gain is retained under a score that matches the derivation.
  3. [§5.3, Tables 4–5; §4.1.2 Eq. (4)] Tables 4–5 isolate w/o U-Bias, w/o I-Bias, w/o Agg., and w/o CL, but do not fully separate “bias features as gates” from “backdoor-style intervention.” A controlled variant that keeps the same bias features and MoE/CL machinery while removing the multiplicative debiased base-score construction (Eq. 4) would show whether the causal parameterization itself, rather than extra capacity and bias covariates, is load-bearing for the gains claimed in §5.2.
minor comments (5)
  1. [§4.1.1 / Appendix A] Main text refers to “Detailed derivation of Eq. 14” while the interventional formula is numbered Eq. (2) in §4.1.1; align equation numbers between body and Appendix A.
  2. [Table 2] Table 2: BCIPM+MCLMR improvements lack the paired t-test significance markers that appear for CRGCN+MCLMR and HEC-GCN+MCLMR; either add tests or explain the omission.
  3. [Figure 1] Figure 1 is dense (many overlapping symbols and paths). A cleaner schematic separating causal estimation, MoE aggregation, and contrastive alignment would help readers map modules to sections.
  4. [§5.1.4 / Appendix B] Hyperparameter search spaces are deferred to Appendix B; a short main-text note on the final (γ_u, γ_i, τ_0, α, expert dim) used for the reported tables would aid reproducibility without opening the appendix.
  5. [§2.2] Related work could more sharply distinguish MCLMR from CMSR and CVID on the dual user/item multi-behavior confounder modeling claim, since both are already listed as integrated causal multi-behavior baselines.

Circularity Check

0 steps flagged

No load-bearing circularity: backdoor adjustment is standard Pearl, relative-frequency proxies are explicit modeling choices (not definitional), and gains are measured on held-out HR/NDCG against independent baselines and ablations.

full rationale

The derivation chain begins from a standard causal graph and Pearl backdoor formula (Eq. 2 / Appendix A), which is external and not self-referential. Bias terms b_u,k and b_i,k are defined as observed relative frequencies (Eq. 3) and then inserted as proxies into the parameterized interventional scores (Eqs. 4–5) and MoE/CL gates; this is a conventional approximation (explicitly flagged by the authors as non-optimal and inherited from DecRS/PDA-style work), not a self-definitional loop or a fit that forces the target prediction. The Adaptive Aggregation and Bias-aware CL modules are architectural designs whose temperature schedule τ=τ0(1−α·b) is a fixed functional form of the same proxy, not a quantity fitted to and then re-predicted from the evaluation metric. Final ranking uses plain inner-product embeddings shaped by the training objective; the training–inference gap is a design choice, not circularity. Empirical claims rest on external multi-behavior and pluggable-causal baselines (Tables 2–3) plus component ablations (Tables 4–5), which supply independent evidence. No uniqueness theorem, self-citation chain, or ansatz smuggled via overlapping authors is load-bearing for the central result. The only minor softness is the untested fidelity of the frequency proxies themselves, which is an assumption-risk issue rather than circular reduction of prediction to input.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard causal adjustment plus several modeling choices and many tuned scalars. The load-bearing modeling axioms are that frequency ratios proxy latent multi-behavior confounders and that intervening via the parameterized product form yields better target-behavior rankings. Free parameters (γ_u, γ_i, MoE dim, τ0, α, λ_J, λ_M, β, CL weight) are grid-searched per dataset/backbone. Invented entities are architectural modules rather than new physical objects; they have no independent evidence outside the reported ablations.

free parameters (5)
  • debiasing coefficients γ_u, γ_i = often 0.01 for items; user γ varies by dataset
    Scale user and item multi-behavior bias terms; searched in {0.01, 0.1, 0.5, 1.0}; optimal values are dataset-dependent (Fig. 3).
  • MoE expert hidden dimension = 64
    Capacity of the semantic gate; tuned in {32, 64, 128}; peaks at 64 (Fig. 4).
  • base temperature τ0 and bias sensitivity α
    Control bias-aware InfoNCE temperature τ=τ0(1−α·b); α∈[0,1); compared against fixed/learnable/random variants in Appendix C.
  • path weights λ_J, λ_M and CL balance β / λ_CL
    Balance Jaccard vs MoE contribution and contrastive vs BPR objectives; learnable or tuned.
  • embedding dimension d = 64
    Fixed at 64 for fair comparison; still a capacity choice that affects all reported metrics.
axioms (5)
  • ad hoc to paper User and item multi-behavior relative frequencies are valid proxies for latent confounders B_u and B_i in the multi-behavior causal graph.
    Stated in §4.1.2 with acknowledgment that optimal proxies are an open problem; follows conventions of DecRS/iDCF/PDA-style work.
  • standard math Backdoor adjustment via do(U), do(I) blocks confounding paths U←B_u→M→Y and I←B_i→M→Y (Fig. 2, Eq. 2 / Appendix A).
    Standard Pearl do-calculus under the assumed DAG; derivation in Appendix A.
  • domain assumption P(M_k=1|·) ≈ f_k(u,i)·g_k(b_u,k,b_i,k) and outcome probabilities factor similarly (Eqs. 4–5).
    Multiplicative parameterization chosen for tractability; not derived from a likelihood model of the data-generating process.
  • domain assumption Auxiliary behaviors have cascading sequential structure (view→cart→buy) that justifies using downstream bias for temperature and contribution gates.
    Used in §4.3.2 and problem setup; holds for the three e-commerce datasets but is not universal for all multi-behavior settings.
  • ad hoc to paper At inference, inner product of target-behavior embeddings is an unbiased preference estimate after causal training (Eq. 13, §4.4).
    Training optimizes S_final with bias and aggregation terms; inference drops them. Assumes training has already removed confounding from the embeddings.
invented entities (3)
  • Bias proxies b_u,k and b_i,k as ‘holistic conversion’ rates no independent evidence
    purpose: Stand in for latent multi-behavior confounders in the intervention formula and gates.
    Defined in Eq. 3; no external validation that they equal true propensities beyond improved ranking metrics.
  • Dual-path Adaptive Aggregation (Jaccard structural gate + MoE semantic gate) no independent evidence
    purpose: Dynamically fuse heterogeneous auxiliary behaviors while limiting negative transfer.
    Architectural construct; support is ablation-only (w/o Jaccard, w/o MoE, w/o Agg.).
  • Bias-aware personalized temperature for multi-behavior InfoNCE no independent evidence
    purpose: Align cross-behavior representations more strictly for high-bias entities and more softly for sparse ones.
    Eq. 10; variants in Appendix C show gains but no external theory of optimal τ(b).

pith-pipeline@v1.1.0-grok45 · 34460 in / 4123 out tokens · 37342 ms · 2026-07-13T18:25:23.920356+00:00 · methodology

0 comments
read the original abstract

Multi-Behavior Recommendation (MBR) leverages multiple user interaction types (e.g., views, clicks, purchases) to enrich preference modeling and alleviate data sparsity issues in traditional single-behavior approaches. However, existing MBR methods face fundamental challenges: they lack principled frameworks to model complex confounding effects from user behavioral habits and item multi-behavior distributions, struggle with effective aggregation of heterogeneous auxiliary behaviors, and fail to align behavioral representations across semantic gaps while accounting for bias distortions. To address these limitations, we propose MCLMR, a novel model-agnostic causal learning framework that can be seamlessly integrated into various MBR architectures. MCLMR first constructs a causal graph to model confounding effects and performs interventions for unbiased preference estimation. Under this causal framework, it employs an Adaptive Aggregation module based on Mixture-of-Experts to dynamically fuse auxiliary behavior information and a Bias-aware Contrastive Learning module to align cross-behavior representations in a bias-aware manner. Extensive experiments on three real-world datasets demonstrate that MCLMR achieves significant performance improvements across various baseline models, validating its effectiveness and generality. All data and code will be made publicly available. For anonymous review, our code is available at the following the link: https://github.com/gitrxh/MCLMR.

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

56 extracted references · 6 canonical work pages

  1. [1]

    Himan Abdollahpouri and Masoud Mansoury. 2020. Multi-sided Exposure Bias in Recommendation.arXiv e-prints, Article arXiv:2006.15772 (June 2020), arXiv:2006.15772 pages. arXiv:2006.15772 [cs.IR] doi:10.48550/arXiv.2006.15772

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.15772 2020

  2. [2]

    Stephen Bonner and Flavian Vasile. 2018. Causal embeddings for recommendation. InProceedings of the 12th ACM Conference on Recommender Systems(Vancouver, British Columbia, Canada)(RecSys ’18). Association for Computing Machinery, New York, NY, USA, 104–112. doi:10.1145/3240323.3240360

    doi:10.1145/3240323.3240360 2018

  3. [3]

    Charles, D

    Léon Bottou, Jonas Peters, Joaquin Quiñonero Candela, Denis X. Charles, D. Max Chickering, Elon Portugaly, Dipankar Ray, Patrice Simard, and Ed Snelson. 2013. Counterfactual reasoning and learning systems: the example of computational advertising.J. Mach. Learn. Res.14, 1 (Jan. 2013), 3207–3260

    2013

  4. [4]

    Wei Cai, ZhiHong Zheng, Xuan Zhang, Weiyi Shang, Yubin Ma, WenJie Gao, and Zhi Jin. 2025. Neighborhood structure enhancement and denoising method for multi-behavior recommendation.Neural Networks191 (2025), 107760. doi:10. 1016/j.neunet.2025.107760

    arXiv 2025

  5. [6]

    arXiv:2010.03240 [cs.IR] https://arxiv.org/abs/2010.03240

    Bias and Debias in Recommender System: A Survey and Future Directions. arXiv:2010.03240 [cs.IR] https://arxiv.org/abs/2010.03240

    Pith/arXiv arXiv 2010

  6. [7]

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

  7. [8]

    ACM Trans

    Bias and Debias in Recommender System: A Survey and Future Directions. ACM Trans. Inf. Syst.41, 3, Article 67 (Feb. 2023), 39 pages. doi:10.1145/3564284

    doi:10.1145/3564284 2023

  8. [9]

    Yuzhe Chen, Jie Cao, Youquan Wang, Jia Wu, Huanhuan Chen, and Guandong Xu. 2025. Causal Variational Inference for Deconfounded Multi-Behavior Rec- ommendation.ACM Trans. Inf. Syst.43, 6, Article 151 (Sept. 2025), 26 pages. doi:10.1145/3745023

    work page doi:10.1145/3745023 2025

  9. [10]

    Chen Gao, Xiangnan He, Dahua Gan, Xiangning Chen, Fuli Feng, Yong Li, Tat- Seng Chua, and Depeng Jin. 2019. Neural multi-task recommendation from multi-behavior data. InICDE. IEEE, 1554–1557

    2019

  10. [11]

    Shuyun Gu, Xiao Wang, Chuan Shi, and Ding Xiao. 2022. Self-supervised Graph Neural Networks for Multi-behavior Recommendation. InProceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI-22. 2052– 2058

    2022

  11. [12]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation.arXiv preprint arXiv:2002.02126(2020)

    Pith/arXiv arXiv 2020

  12. [13]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InWWW. 173–182

    2017

  13. [14]

    Chao Huang. 2021. Recent Advances in Heterogeneous Relation Learning for Recommendation.arXiv preprint arXiv:2110.03455(2021)

    Pith/arXiv arXiv 2021

  14. [15]

    Bowen Jin, Chen Gao, Xiangnan He, Depeng Jin, and Yong Li. 2020. Multi- behavior recommendation with graph convolutional networks. InSIGIR

    2020

  15. [16]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

    Pith/arXiv arXiv 2014

  16. [17]

    Karl Krauth, Yixin Wang, and Michael I. Jordan. 2022. Breaking Feedback Loops in Recommender Systems with Causal Inference. arXiv:2207.01616 [cs.IR] https: //arxiv.org/abs/2207.01616

    Pith/arXiv arXiv 2022

  17. [18]

    Jie Liao, Min Yang, Wei Zhou, Hongyu Zhang, and Junhao Wen. 2024. Modeling item exposure and user satisfaction for debiased recommendation with causal inference.Inf. Sci.676, C (Aug. 2024), 15 pages. doi:10.1016/j.ins.2024.120834

    work page doi:10.1016/j.ins.2024.120834 2024

  18. [19]

    Lingfeng Liu, Yixin Song, Dazhong Shen, Bing Yin, Hao Li, Yanyong Zhang, and Chao Wang. 2025. Rethinking Popularity Bias in Collaborative Filtering via Analytical Vector Decomposition.arXiv preprint arXiv:2512.10688(2025)

    arXiv 2025

  19. [20]

    Dan Lu, Shiqing Wu, Hao Zhang, Guandong Xu, and Qilong Han. 2025. Causal cascading convolution networks for multi-behavior sequential recommendation. Information Sciences720 (2025), 122484. doi:10.1016/j.ins.2025.122484

    work page doi:10.1016/j.ins.2025.122484 2025

  20. [21]

    Huishi Luo, Fuzhen Zhuang, Ruobing Xie, Hengshu Zhu, Deqing Wang, Zhulin An, and Yongjun Xu. 2024. A survey on causal inference for recommendation. The Innovation5, 2 (2024), 100590. doi:10.1016/j.xinn.2024.100590

    doi:10.1016/j.xinn.2024.100590 2024

  21. [22]

    Chenglong Ma, Ziqi Xu, Yongli Ren, Danula Hettiachchi, and Jeffrey Chan. 2025. PUB: An LLM-Enhanced Personality-Driven User Behaviour Simulator for Rec- ommender System Evaluation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR. 2690–

    2025

  22. [23]

    https://doi.org/10.1145/3726302.3730238

    doi:10.1145/3726302.3730238

  23. [24]

    Masoud Mansoury, Bamshad Mobasher, and Herke van Hoof. 2024. Mitigat- ing Exposure Bias in Online Learning to Rank Recommendation: A Novel Re- ward Model for Cascading Bandits. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management(Boise, ID, USA)(CIKM ’24). Association for Computing Machinery, New York, NY, USA, 16...

    doi:10.1145/3627673.3679763 2024

  24. [25]

    2009.Causality: Models, Reasoning and Inference(2nd ed.)

    Judea Pearl. 2009.Causality: Models, Reasoning and Inference(2nd ed.). Cambridge University Press, USA

    2009

  25. [26]

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

  26. [27]

    BPR: Bayesian personalized ranking from implicit feedback.arXiv preprint arXiv:1205.2618(2012)

    Pith/arXiv arXiv 2012

  27. [28]

    Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and Max Welling. 2018. Modeling relational data with graph convolutional networks. InEuropean semantic web conference. Springer, 593–607

    2018

  28. [29]

    Xiaoyuan Su and Taghi M Khoshgoftaar. 2009. A survey of collaborative filtering techniques.Advances in artificial intelligence2009 (2009)

    2009

  29. [30]

    Adith Swaminathan and Thorsten Joachims. 2015. The Self-Normalized Estimator for Counterfactual Learning. InAdvances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett (Eds.), Vol. 28. Curran Associates, Inc. https://proceedings.neurips.cc/paper_files/paper/2015/ file/39027dfad5138c9ca0c474d71db915c3-Paper.pdf

    2015

  30. [31]

    Chao Wang, Qi Liu, Runze Wu, Enhong Chen, Chuanren Liu, Xunpeng Huang, and Zhenya Huang. 2018. Confidence-aware matrix factorization for recom- mender systems. InProceedings of the AAAI Conference on artificial intelligence, Vol. 32

    2018

  31. [32]

    Chao Wang, Yixin Song, Jinhui Ye, Chuan Qin, Dazhong Shen, Lingfeng Liu, Xiang Wang, and Yanyong Zhang. 2025. FACE: A General Framework for Mapping Col- laborative Filtering Embeddings into LLM Tokens.arXiv preprint arXiv:2510.15729 (2025)

    arXiv 2025

  32. [33]

    Chao Wang, Hengshu Zhu, Qiming Hao, Keli Xiao, and Hui Xiong. 2021. Vari- able interval time sequence modeling for career trajectory prediction: Deep collaborative perspective. InProceedings of the Web Conference 2021. 612–623

    2021

  33. [34]

    Chao Wang, Hengshu Zhu, Peng Wang, Chen Zhu, Xi Zhang, Enhong Chen, and Hui Xiong. 2021. Personalized and explainable employee training course recom- mendations: A bayesian variational approach.ACM Transactions on Information Systems (TOIS)40, 4 (2021), 1–32

    2021

  34. [35]

    Chao Wang, Hengshu Zhu, Chen Zhu, Chuan Qin, and Hui Xiong. 2020. Setrank: A setwise bayesian approach for collaborative ranking from implicit feedback. In Proceedings of the aaai conference on artificial intelligence, Vol. 34. 6127–6136

    2020

  35. [36]

    Yixin Wang, Dawen Liang, Laurent Charlin, and David M. Blei. 2018. The De- confounded Recommender: A Causal Inference Approach to Recommendation. arXiv e-prints, Article arXiv:1808.06581 (Aug. 2018), arXiv:1808.06581 pages. arXiv:1808.06581 [cs.IR] doi:10.48550/arXiv.1808.06581

    Pith/arXiv arXiv doi:10.48550/arxiv.1808.06581 2018

  36. [37]

    Xinye Wanyan, Danula Hettiachchi, Chenglong Ma, Ziqi Xu, and Jeffrey Chan

  37. [38]

    InProceedings of the 34th ACM International Con- ference on Information and Knowledge Management, CIKM

    Temporal-Aware User Behaviour Simulation with Large Language Models for Recommender Systems. InProceedings of the 34th ACM International Con- ference on Information and Knowledge Management, CIKM. 5335–5339. https: //doi.org/10.1145/3746252.3760878

    doi:10.1145/3746252.3760878

  38. [39]

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

  39. [40]

    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. 1791–1800

  40. [41]

    Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Mengyin Lu, and Liefeng Bo. 2021. Multi-Behavior Enhanced Recommendation with Cross-Interaction Collaborative Relation Modeling. InICDE. IEEE, 1931–1936

    2021

  41. [42]

    Lianghao Xia, Chao Huang, Yong Xu, Peng Dai, Bo Zhang, and Liefeng Bo

  42. [43]

    Multiplex behavioral relation learning for recommendation via memory augmented transformer network. InSIGIR. 2397–2406

  43. [44]

    Ziqi Xu, Debo Cheng, Jiuyong Li, Jixue Liu, Lin Liu, and Ke Wang. 2023. Disen- tangled Representation for Causal Mediation Analysis. InThirty-Seventh AAAI Conference on Artificial Intelligence, AAAI. 10666–10674. https://doi.org/10.1609/ aaai.v37i9.26266

    2023

  44. [45]

    Ziqi Xu, Debo Cheng, Jiuyong Li, Jixue Liu, Lin Liu, and Kui Yu. 2024. Causal Inference with Conditional Front-Door Adjustment and Identifiable Variational Autoencoder. InThe Twelfth International Conference on Learning Representations, ICLR. https://openreview.net/forum?id=wFf9m4v7oC

    2024

  45. [46]

    Hongrui Xuan, Yi Liu, Bohan Li, and Hongzhi Yin. 2023. Knowledge Enhancement for Contrastive Multi-Behavior Recommendation. InProceedings of the Sixteenth ACM International Conference on Web Search and Data Mining. 195–203

    2023

  46. [47]

    Mingshi Yan, Zhiyong Cheng, Chen Gao, Jing Sun, Fan Liu, Fuming Sun, and Haojie Li. 2023. Cascading Residual Graph Convolutional Network for Multi- Behavior Recommendation. 42, 1, Article 10 (Aug. 2023), 26 pages. doi:10.1145/ 3587693

    2023

  47. [48]

    Mingshi Yan, Zhiyong Cheng, Jing Sun, Fuming Sun, and Yuxin Peng. 2023. MB-HGCN: A hierarchical graph convolutional network for multi-behavior rec- ommendation.arXiv preprint arXiv:2306.10679(2023)

    Pith/arXiv arXiv 2023

  48. [49]

    Mingshi Yan, Fan Liu, Jing Sun, Fuming Sun, Zhiyong Cheng, and Yahong Han

  49. [50]

    InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Behavior-Contextualized Item Preference Modeling for Multi-Behavior Recommendation. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 946–955

  50. [51]

    Yabo Yin, Xiaofei Zhu, Wenshan Wang, Yihao Zhang, Pengfei Wang, Yixing Fan, and Jiafeng Guo. 2024. HEC-GCN: Hypergraph Enhanced Cascading Graph Con- volution Network for Multi-Behavior Recommendation. arXiv:2412.14476 [cs.IR] https://arxiv.org/abs/2412.14476

    Pith/arXiv arXiv 2024

  51. [52]

    Guixian Zhang, Guan Yuan, Debo Cheng, Lin Liu, Jiuyong Li, Ziqi Xu, and Shichao Zhang. 2025. Deconfounding representation learning for mitigating latent confounding effects in recommendation.Knowledge and Information Systems67, 7 (2025), 5999–6020. https://doi.org/10.1007/s10115-025-02404-7

    work page doi:10.1007/s10115-025-02404-7 2025

  52. [53]

    Wenhao Zhang, Wentian Bao, Xiao-Yang Liu, Keping Yang, Quan Lin, Hong Wen, and Ramin Ramezani. 2019. A Causal Perspective to Unbiased Conversion Rate Estimation on Data Missing Not at Random.CoRRabs/1910.09337 (2019). arXiv:1910.09337 http://arxiv.org/abs/1910.09337 WWW ’26, April 13–17, 2026, Dubai, United Arab Emirates Ranxu Zhang et al

    Pith/arXiv arXiv 2019

  53. [54]

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. 2021. Causal Intervention for Leveraging Popularity Bias in Recommendation. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval(Virtual Event, Canada)(SIGIR ’21). Association for Computing Machiner...

    work page doi:10.1145/3404835.3462875 2021

  54. [55]

    Weiqi Zhao, Dian Tang, Xin Chen, Dawei Lv, Daoli Ou, Biao Li, Peng Jiang, and Kun Gai. 2023. Disentangled Causal Embedding With Contrastive Learning For Recommender System. InCompanion Proceedings of the ACM Web Conference 2023(Austin, TX, USA)(WWW ’23 Companion). Association for Computing Machinery, New York, NY, USA, 406–410. doi:10.1145/3543873.3584637

    work page doi:10.1145/3543873.3584637 2023

  55. [56]

    Zhe Zhao, Zhiyuan Cheng, Lichan Hong, and Ed H Chi. 2015. Improving user topic interest profiles by behavior factorization. InProceedings of the 24th International Conference on World Wide Web. 1406–1416

    2015

  56. [57]

    Yunqin Zhu, Chao Wang, Qi Zhang, and Hui Xiong. 2024. Graph signal diffusion model for collaborative filtering. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1380– 1390. A Causal Effect Derivation To block the backdoor paths created by confounding biases, we apply causal intervention usi...

    arXiv 2024