Pith. sign in

REVIEW 2 major objections 4 minor 43 references

Counterfactual Learning-Driven Representation Disentanglement for Search-Enhanced Recommendation

T0 review · 2 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Counterfactual score shifts supervise a split of search-clicked item features into query-specific and general-interest parts, improving recommendation.

desk verdict The paper's core counterfactual triplet loss appears to invert its own inequalities, so the claimed disentanglement mechanism likely does not do what is described. read the letter →

arxiv 2411.18631 v1 pith:RXS6XFKZ submitted 2024-11-14 cs.IR

Jiajun Cui , Xu Chen , Shuai Xiao , Chen Ju , Jinsong Lan , Qingwen Liu , Wei Zhang This is my paper · ORCID
classification cs.IR
keywords search-enhancedrecommendationcounterfactuallearningrepresentationdisentanglementnegativetransferrecommendersystemsfeatureaugmentationdatasequential
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

Search-enhanced recommenders use a user's search clicks to learn preferences, but those clicks mix features that match the query being searched with features that reflect the user's general taste. The paper argues that transferring both kinds of features into the recommendation domain leaks search-specific intent and can hurt recommendation accuracy. ClardRec constructs counterfactual supervision from score changes: if removing query-related features changes query-item scores more than user-item scores, and removing general-interest features changes user-item scores more, those relative changes can teach an attribute-aware gating network to split item representations. The query-independent parts are then transferred via feature and data augmentation. Experiments on a short-video dataset and an industrial e-commerce dataset report accuracy gains across collaborative filtering and sequential recommendation backbones.

What carries the argument

The load-bearing mechanism is the triplet counterfactual objective built on score-variation ordering. The model computes user-item preference and query-item match scores, $\hat{s}_{ui}=u^{\mathrm{src}}\cdot i^{\mathrm{src}}$ and $\hat{s}_{qi}=q^{\mathrm{src}}\cdot i^{\mathrm{src}}$, and the corresponding scores $\hat{s}_{ui/q},\hat{s}_{qi/q},\hat{s}_{ui/u},\hat{s}_{qi/u}$ after replacing $i^{\mathrm{src}}$ with the disentangled representations $i^{\mathrm{src}}_{/q}$ and $i^{\mathrm{src}}_{/u}$. The supervision is the assumed ordering $|\Delta_{qi/q}|>|\Delta_{ui/q}|$ and $|\Delta_{ui/u}|>|\Delta_{qi/u}|$, together with $|\Delta_{qi/q}|>|\Delta_{qi/u}|$ and $|\Delta_{ui/u}|>|\Delta_{ui/q}|$, where each $\Delta$ is the corresponding score difference (e.g., $\Delta_{qi/q}=\hat{s}_{qi/q}-\hat{s}_{qi}$). These inequalities become log-sigmoid triplet losses and constraint losses that train an attribute-aware gated network to split each search item embedding into a query-independent part and a query-related part. The query-independent part is then transferred through a gated fusion network for feature augmentation and a confidence-weighted cross-entropy loss for data augmentation.

What would settle it

On a dataset where item attributes are labeled as query-specific or general-interest, train ClardRec and inspect whether the query-independent representation $i^{\mathrm{src}}_{/q}$ preferentially receives the labeled general-interest attributes; the central claim is falsified if those attributes instead land in $i^{\mathrm{src}}_{/u}$, or if replacing the counterfactual ordering losses with shuffled score-orderings leaves recommendation accuracy unchanged.

Watch

Extended reading notes

Core claim

The paper's central claim is that the ordering of score changes under counterfactual feature removal is a usable supervision signal for disentangling item representations without feature-level labels. For a clicked (user, query, item) triple, removing the query-related part of the item (leaving the query-independent representation $i^{\mathrm{src}}_{/q}$) should change the query-item match score more than the user-item preference score, while removing the general-interest part (leaving the query-related representation $i^{\mathrm{src}}_{/u}$) should change the user-item preference score more. These matching-wise inequalities are paired with removal-wise inequalities comparing the two removal operations, and all are encoded as log-sigmoid triplet losses and score-constraint losses. The same attribute-aware gated network that produces the split also yields gate weights per item attribute, so attributes like category can carry different tendencies for query matching versus general interest. The paper claims that the query-independent representation obtained this way reflects user general interest and that transferring it through feature and data augmentation improves recommendation accuracy.

Load-bearing premise

The load-bearing premise is that the relative sizes of score changes—query-item scores moving more when query-related features are removed, user-item scores moving more when general-interest features are removed—correctly identify which item features are search-specific and which are general taste, since the learned split is never checked against ground-truth feature labels.

Editorial extensions

If this is right

  • Platforms with overlapping search and recommendation identities can obtain general-interest item representations without hand-labeled feature annotations.
  • The same feature augmentation can be grafted onto different collaborative filtering and sequential backbones, so the benefit is not tied to one architecture.
  • Confidence-weighted data augmentation down-weights search clicks that mostly reflect query intent, which is the mechanism that mitigates negative transfer from search data.
  • Because the method requires overlapping users between the two domains, its gains in the reported settings are conditional on that overlap; the paper names this as its main limitation.

Reading between the lines

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

  • Editorial inference: the score-variation recipe does not depend on search specifically; the same inequalities could treat any contextual variable (session, campaign, device) as the 'query', yielding a generic way to separate context-bound intent from stable preference.
  • Editorial inference: a direct validity check the paper does not report is attribute-level inspection—comparing gate weights for known search-specific versus general attributes—which would show whether the counterfactual inequalities identify semantically meaningful features or just score sensitivity.
  • Editorial inference: transfer could go both directions; the disentangled representations learned for recommendation items could in turn be used to explain search results, surfacing the general-interest versus query-intent weights behind each match.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes ClardRec, a search-enhanced recommendation framework that uses search queries to disentangle item representations into query-related and query-independent components. The query-independent component is then transferred to the recommendation domain through feature augmentation (a gated fusion network) and data augmentation (confidence-weighted contrastive/auxiliary losses). The method is evaluated on KuaiSAR and an industrial E-commerce R&S dataset under both collaborative filtering and sequential recommendation backbones, with consistent reported improvements over baselines.

Significance. If the disentanglement mechanism worked as claimed, the paper would be a useful practical contribution to search-enhanced recommendation, addressing a real negative-transfer problem. The empirical protocol has strengths: recommendation metrics are held out, 99 negative items are used, five random seeds are reported, and multiple backbones and recent baselines are considered. However, the central supervision signal has a sign inconsistency that inverts the stated score-variation inequalities, and the disentanglement is not validated against any external ground truth. Until those issues are resolved, the reported gains cannot be attributed to the claimed counterfactual disentanglement mechanism.

major comments (2)
  1. [Section 4.2.2, Eqs. (10)-(14)] With Delta defined as after-before in Section 4.2.2, the triplet losses in Eqs. (13) and (14) reward the opposite of the inequalities in Eqs. (10)-(12). For a positive item, Eq. (10) states |Delta_q i/q| > |Delta_u i/q|; combined with the paper's own sign statement Delta_q i/q <= 0 and Delta_u i/q <= 0, this is equivalent to Delta_q i/q - Delta_u i/q < 0. The first term of L/q_cd, -log sigma(Delta_q i/q - Delta_u i/q), is minimized by making Delta_q i/q - Delta_u i/q as large as possible, i.e., Delta_q i/q > Delta_u i/q. The negative-item term is inverted in the same way: Eq. (10) for a negative item implies Delta_q i-/q > Delta_u i-/q, while the loss maximizes Delta_u i-/q - Delta_q i-/q. The same reversal affects L/u_cd in Eq. (13) and both terms of Eq. (14) against Eq. (12). Consequently, L_cd as written does not implement the stated counterfactual supervision, and the query-independent representation i_src/q is not trained to satisfy the assumed ordering. The ablation results in Table 2 therefore cannot be attributed to the claimed disentanglement mechanism.
  2. [Section 4.2.2, Eqs. (10)-(12) and Section 5.4] The disentanglement is supervised only by the assumed score-variation orderings; there is no external check that i_src/q actually encodes query-independent, recommendation-relevant item semantics. The t-SNE visualization in Figure 3 and the case study in Figure 4 are qualitative illustrations, not quantitative measurements. I recommend adding a probe test, such as a classifier predicting query words or item categories from i_src/q versus i_src/u, or an ablation in which the gated split is replaced by a random split with a random score-ordering objective, to demonstrate that the gains come from genuine disentanglement rather than from added model capacity or generic regularization.
minor comments (4)
  1. [Section 4.2.1] The notation i_emb/q and i_emb/u is confusing: i_emb/q is defined as the embedding after removing query-related features, yet the '/q' subscript more naturally suggests query-related content. Please rename the two components (e.g., i_gen and i_qry) and state the convention explicitly at first use.
  2. [Section 5.1.3 and Table 1] The asterisks are declared to indicate statistical significance over the best baseline, but several bold entries (e.g., MLP+ClardRec on KuaiSAR HR@1) are not starred; please clarify when differences are significant and specify the number of seeds and the exact test procedure.
  3. [Section 5.4 and Figure 3] The t-SNE visualization is presented as evidence of successful disentanglement, but no quantitative separation or alignment measure is reported; adding a distance- or centroid-based metric would make the answer to Q3 more convincing.
  4. [References] Reference [31] appears garbled ('K. Z. Zhang P.K. Kannan Y Yang'); please correct the author list.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main effectiveness claim is anchored by held-out recommendation metrics, and the counterfactual objective is a training constraint rather than a derived prediction. An internal sign inconsistency in the triplet losses is noted as a correctness risk, not as circularity.

full rationale

ClardRec's headline claim, that it improves recommendation in collaborative filtering and sequential scenarios, is evaluated on held-out HR@1/HR@5/NDCG@5/MRR against baselines; the recommendation loss L_rec and the transferred representations are tested on data not used to fit the gates, fusion network, or loss weights. The counterfactual disentangling module is presented as a supervision scheme: the paper says it 'propose[s] triplet counterfactual objectives to explicitly supervise the learning of query-independent and query-related item representations' (Sec. 4.2.2), and the score-change inequalities in Eqs. 10-12 are imposed on the model through L_cd rather than being predicted from it. Thus the statement 'Using this paradigm, we can obtain the item representation i_src/q, which is rich in query-independent general features' is a description of the training objective, not a prediction that is then verified by the same objective. Self-citations in the related work (e.g., [4], [7], [29]) are contextual and are not load-bearing for the central derivation; no uniqueness theorem is imported from the authors' prior work. The internal validation in Sec. 5.4 does rely on the model's own representations, and the confidence weights in Eq. 21 come from the same model's ŵs_{u i/q}; this is weak, self-referential evidence for the semantic claim of disentanglement, but it does not make the held-out recommendation result circular. Separately, as a correctness issue rather than circularity: under the paper's sign convention Δ = after − before, the log-sigmoid triplet losses in Eqs. 13-14 appear to maximize the opposite ordering of the score variations required by Eqs. 10-12 for positive and negative items (e.g., for a positive item Eq. 10 requires Δ_{q i/q} < Δ_{u i/q}, while −logσ(Δ_{q i/q} − Δ_{u i/q}) is minimized by making Δ_{q i/q} > Δ_{u i/q}). This internal inconsistency undermines the claim that L_cd implements the stated counterfactual signals, but it is a derivation bug, not a circular reduction of the paper's conclusion to its inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim relies on the semantic decomposition of item embeddings into query-related and query-independent parts. This decomposition is implemented by learned gates and supervised only by assumed inequalities, not by independent labels. The loss-balance hyperparameters are tuned per dataset and backbone. No new physical entity is introduced; the latent disentangled representations are internal constructs whose verification is indirect. This is a moderate assumption load for an applied recommendation method.

free parameters (4)
  • Loss-balance factors lambda, alpha, beta, gamma = Per dataset and backbone; e.g., KuaiSAR MLP: 0.1, 0.1, 0.2, 0.5; E-commerce SASRec: 0.2, 0.1, 1.0, 1.0
    Introduced in Eq. 23 and tuned from {0.01, 0.1, 0.2, 0.5, 1.0} in Table 4. The reported performance peaks depend on these chosen values.
  • Gate and fusion network parameters = Learned; final values not reported
    The z_ID and z_a gates in Eqs. 6 to 8 define the disentanglement split, and the fusion gate in Eq. 19 weights the transfer. These are fitted by gradient descent without external disentanglement labels.
  • Backbone depth L and item encoder depth L_e = L in {1,2,3}, L_e in {1,3} per Table 4
    Architecture capacity is tuned per dataset and backbone, which affects the reported gains.
  • Learning rate and batch size = See Table 4; e.g., 0.005/4096 for MLP on KuaiSAR
    Standard training hyperparameters selected per dataset and backbone; they influence all compared methods equally in principle.
assumptions (5)
  • domain assumption A click under a query depends on both item-query relevance and query-independent item features that interest the user.
    Used in Sections 1 and 4.2.2 as the motivating belief (Figure 1c). No direct labels distinguish the two feature types.
  • ad hoc to paper Score-variation orderings in Eqs. 10 to 12 hold for the learned representations.
    These inequalities are the sole supervision for disentanglement. They are asserted, not derived or independently validated.
  • ad hoc to paper The gated decomposition i_emb = (i_emb/q + i_emb/u)/2 exactly separates query-related from query-independent semantics.
    Eq. 6 defines the split with gates, but no identifiability or orthogonality condition guarantees the parts correspond to the intended concepts.
  • domain assumption Randomly sampled unclicked items are valid negatives.
    Used in Eqs. 4, 13, and 14 for training and evaluation; some negatives may be false negatives.
  • domain assumption Sufficient user and item overlap exists between search and recommendation domains.
    Section 4.1 shares bottom embeddings across domains, and Appendix A concedes the method does not handle non-overlapping users.
invented entities (1)
  • Latent disentangled item representations i_src/q and i_src/u
    purpose: To isolate query-independent general features for transfer to recommendation
    These internal latent variables are produced by learned gates (Eqs. 6 to 9). There is no external measurement or ground-truth label confirming they correspond to query-independent versus query-related features; evidence is indirect (metrics, t-SNE, one case study).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Counterfactual Learning-Driven Representation Disentanglement for Search-Enhanced Recommendation." pith.science (2026). https://pith.science/paper/RXS6XFKZ

@misc{pith2026241118631,
  author       = {Pith},
  title        = {Pith review of: Counterfactual Learning-Driven Representation Disentanglement for Search-Enhanced Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RXS6XFKZ}},
  note         = {Machine review of arXiv:2411.18631}
}
read the original abstract

For recommender systems in internet platforms, search activities provide additional insights into user interest through query-click interactions with items, and are thus widely used for enhancing personalized recommendation. However, these interacted items not only have transferable features matching users' interest helpful for the recommendation domain, but also have features related to users' unique intents in the search domain. Such domain gap of item features is neglected by most current search-enhanced recommendation methods. They directly incorporate these search behaviors into recommendation, and thus introduce partial negative transfer. To address this, we propose a Counterfactual learning-driven representation disentanglement framework for search-enhanced recommendation, based on the common belief that a user would click an item under a query not solely because of the item-query match but also due to the item's query-independent general features (e.g., color or style) that interest the user. These general features exclude the reflection of search-specific intents contained in queries, ensuring a pure match to users' underlying interest to complement recommendation. According to counterfactual thinking, how would user preferences and query match change for items if we removed their query-related features in search, we leverage search queries to construct counterfactual signals to disentangle item representations, isolating only query-independent general features. These representations subsequently enable feature augmentation and data augmentation for the recommendation scenario. Comprehensive experiments on real datasets demonstrate ClardRec is effective in both collaborative filtering and sequential recommendation scenarios.

Figures

Figures reproduced from arXiv: 2411.18631 by the authors.

Figure 1
Figure 1. (a) t-SNE visualization of four active users’ different clicked product distributions in search and recommendation. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of ClardRec. “REC” and “SRC” denote the recommendation and search domains. The input of user historical [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. A real case of ClardRec enhancing recommendation [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Hyper-parameter analysis to investigate the effects of counterfactual disentangling and data augmentation in ClardRec. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 32 canonical work pages

  1. [1]

    Ehsan Abbasnejad, Damien Teney, Amin Parvaneh, Javen Shi, and Anton van den Hengel. 2020. Counterfactual Vision and Language Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2020

  2. [2]

    Stephen Bonner and Flavian Vasile. 2018. Causal embeddings for recommendation. In Proceedings of the 12th ACM Conference on Recommender Systems . 104–112

    work page 2018

  3. [3]

    Fei Cai, Shangsong Liang, and Maarten de Rijke. 2014. Personalized document re-ranking based on Bayesian probabilistic matrix factorization. In Proceedings of the 37th International ACM SIGIR Conference on Research & Development in Information Retrieval. 835–838

    work page 2014

  4. [4]

    Xu Chen, Zida Cheng, Jiangchao Yao, Chen Ju, Weilin Huang, Jinsong Lan, Xiaoyi Zeng, and Shuai Xiao. 2024. Enhancing Cross-Domain Click-Through Rate Prediction via Explicit Feature Augmentation. In Companion Proceedings of the ACM on Web Conference 2024 . 423–432

    work page 2024

  5. [5]

    Tsang, Yuangang Pan, and Jingchao Su

    Xu Chen, Ya Zhang, Ivor W. Tsang, Yuangang Pan, and Jingchao Su. 2023. To- ward Equivalent Transformation of User Preferences in Cross Domain Rec- ommendation. ACM Trans. Inf. Syst. 41, 1, Article 14 (jan 2023), 31 pages. https://doi.org/10.1145/3522762

    work page doi:10.1145/3522762 2023

  6. [6]

    Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio

  7. [7]

    Jiajun Cui, Minghe Yu, Bo Jiang, Aimin Zhou, Jianyong Wang, and Wei Zhang

    work page

  8. [8]

    Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee

Show all 43 references
  1. [9]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  2. [10]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. In Proceedings of the 22nd ACM International Conference on Information & Knowledge Management . 2333–2338

    2013

  3. [11]

    Jianchao Ji, Zelong Li, Shuyuan Xu, Max Xiong, Juntao Tan, Yingqiang Ge, Hao Wang, and Yongfeng Zhang. 2023. Counterfactual Collaborative Reasoning. In Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining. 249–257

    2023

  4. [12]

    Meng Jian, Jingjing Guo, Ye Xiang, and Lifang Wu. 2021. Counterfactual Embed- ding Learning for Debiased Recommendation. In 2021 IEEE Seventh International Conference on Multimedia Big Data (BigMM) . 69–73

    2021

  5. [13]

    Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Rec- ommendation. In 2018 IEEE International Conference on Data Mining (ICDM) . 197–206

    2018

  6. [14]

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

    2014 arXiv

  7. [15]

    Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018. Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture- of-Experts. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining . 1930–1939

    2018

  8. [16]

    Alessandro Micarelli, Fabio Gasparetti, Filippo Sciarrone, and Susan Gauch. 2007. Personalized search on the world wide web. In The Adaptive Web: Methods and Strategies of Web Personalization. 195–230

    2007

  9. [17]

    Amin Parvaneh, Ehsan Abbasnejad, Damien Teney, Javen Qinfeng Shi, and An- ton van den Hengel. 2020. Counterfactual Vision-and-Language Navigation: Unravelling the Unseen. In Advances in Neural Information Processing Systems . 5296–5307

    2020

  10. [18]

    Chen Qian, Fuli Feng, Lijie Wen, Chunping Ma, and Pengjun Xie. 2021. Coun- terfactual Inference for Text Classification Debiasing. In Proceedings of the Joint Conference of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Jo...

    2021

  11. [19]

    Qijie Shen, Hong Wen, Jing Zhang, and Qi Rao. 2022. Hierarchically Fusing Long and Short-Term User Interests for Click-Through Rate Prediction in Product Search. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 1767–1776

    2022

  12. [20]

    Teng Shi, Zihua Si, Jun Xu, Xiao Zhang, Xiaoxue Zang, Kai Zheng, Dewei Leng, Yanan Niu, and Yang Song. 2024. UniSAR: Modeling User Transition Behaviors between Search and Recommendation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development ...

    2024

  13. [21]

    Zihua Si, Xueran Han, Xiao Zhang, Jun Xu, Yue Yin, Yang Song, and Ji-Rong Wen. 2022. A Model-Agnostic Causal Learning Framework for Recommendation using Search Data. In Proceedings of the ACM Web Conference 2022 . 224–233

    2022

  14. [22]

    Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Yang Song, Xiaoxue Zang, and Ji-Rong Wen. 2023. Enhancing Recommendation with Search Data in a Causal Learning Manner. ACM Transactions on Information Systems 41, 4, Article 111 (2023), 31 pages

    2023

  15. [23]

    Zihua Si, Zhongxiang Sun, Xiao Zhang, Jun Xu, Xiaoxue Zang, Yang Song, Kun Gai, and Ji-Rong Wen. 2023. When Search Meets Recommendation: Learning Dis- entangled Search Representation for Recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research ...

    2023

  16. [24]

    Yan Song, Shuming Shi, Jing Li, and Haisong Zhang. 2018. Directional Skip- Gram: Explicitly Distinguishing Left and Right Context for Word Embeddings. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human La...

    2018

  17. [25]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Dewei Leng, Yanan Niu, Yang Song, Xiao Zhang, and Jun Xu. 2023. KuaiSAR: A Unified Search And Recommendation Dataset. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management. 5407–5411

    2023

  18. [26]

    Khanh Hiep Tran, Azin Ghazimatin, and Rishiraj Saha Roy. 2021. Counterfactual Explanations for Neural Recommenders. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York, NY, USA, 1627–1631

    2021

  19. [27]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems. 6000–6010

    2017

  20. [28]

    Pei Wang and Nuno Vasconcelos. 2020. SCOUT: Self-Aware Discriminant Coun- terfactual Explanations. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2020

  21. [29]

    Zhenlei Wang, Jingsen Zhang, Hongteng Xu, Xu Chen, Yongfeng Zhang, Wayne Xin Zhao, and Ji-Rong Wen. 2021. Counterfactual Data-Augmented Sequential Recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval...

    2021

  22. [30]

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

  23. [31]

    Kenny, Tin Lok James Ng, K

    Linyi Yang, Eoin M. Kenny, Tin Lok James Ng, K. Z. Zhang P.K. Kannan Y Yang, Barry Smyth, and Ruihai Dong. 2020. Generating Plausible Counterfactual Explanations for Deep Transformers in Financial Text Classification. InCOLING

    2020

  24. [32]

    Jing Yao, Zhicheng Dou, Ruobing Xie, Yanxiong Lu, Zhiping Wang, and Ji-Rong Wen. 2021. USER: A Unified Information Search and Recommendation Model based on Integrated Behavior Sequence. In Proceedings of the 30th ACM Interna- tional Conference on Information & Knowledge Manage...

    2021

  25. [33]

    Bruce Croft

    Hamed Zamani and W. Bruce Croft. 2018. Joint Modeling and Optimization of Search and Recommendation. In Proceedings of the First Biennial Conference on Design of Experimental Search & Information Retrieval Systems . 36–41

    2018

  26. [34]

    Mehdi Zemni, Mickaël Chen, Éloi Zablocki, Hedi Ben-Younes, Patrick Pérez, and Matthieu Cord. 2023. OCTET: Object-aware Counterfactual Explanations. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR

    2023

  27. [35]

    Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. 2021. Causal Intervention for Leveraging Popularity Bias in Recommendation. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Informatio...

    2021

  28. [36]

    Feng Zhu, Chaochao Chen, Yan Wang, Guanfeng Liu, and Xiaolin Zheng. 2019. DTCDR: A Framework for Dual-Target Cross-Domain Recommendation. In Pro- ceedings of the 28th ACM International Conference on Information and Knowledge Management. 1533–1542

    2019

  29. [37]

    Feng Zhu, Yan Wang, Chaochao Chen, Jun Zhou, Longfei Li, and Guanfeng Liu

  30. [43]

    src. ” and “rec

    Cross-Domain Recommendation: Challenges, Progress, and Prospects. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Jiajun Cui 1, Xu Chen2, Shuai Xiao2, Chen Ju2, Jinsong Lan2, Qin...

    2018

  31. [2014]

    In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation

    On the Properties of Neural Machine Translation: Encoder–Decoder Ap- proaches. In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation. 103–111

  32. [2016]

    In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings

    Session-based Recommendations with Recurrent Neural Networks. In 4th International Conference on Learning Representations, ICLR 2016, San Juan, Puerto Rico, May 2-4, 2016, Conference Track Proceedings

    2016

  33. [2019]

    In Proceedings of the 36th International Conference on Machine Learning

    Counterfactual Visual Explanations. In Proceedings of the 36th International Conference on Machine Learning . 2376–2384

  34. [2021]

    1791–1800

    Model-Agnostic Counterfactual Reasoning for Eliminating Popularity Bias in Recommender System. 1791–1800. https://doi.org/10.1145/3447548.3467289

  35. [2024]

    In Proceedings of the 40th IEEE International Conference on Data Engineering

    Interpretable Knowledge Tracing via Response Influence-based Counter- factual Reasoning. In Proceedings of the 40th IEEE International Conference on Data Engineering

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.