Pith. sign in

REVIEW 4 major objections 6 minor 116 references

HiCore: multi-channel hypergraph learning is claimed to curb the Matthew effect in conversational recommendation and set new state-of-the-art results.

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

2026-08-01 14:50 UTC pith:OBQRUP5Y

load-bearing objection The paper is a near-duplicate of the authors' own EMNLP 2024 paper and Table 1 contains an impossible metric value; the SOTA claim cannot be trusted, though the architecture itself has some plausible components. the 4 major comments →

arxiv 2607.18609 v1 pith:OBQRUP5Y submitted 2026-07-21 cs.IR cs.AIcs.HC

Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation

Yongsen Zheng , Ruilin Xu , Guohua Wang , Liang Lin , Kwok-Yan Lam This is my paper
classification cs.IR cs.AIcs.HC
keywords conversational recommendationMatthew effectpopularity biashypergraph learningself-supervised learningmulti-interest modelingrecommendation diversificationlong-tail items
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 argues that the Matthew effect—where popular items keep getting recommended while niche items are ignored—intensifies in conversational recommender systems because users and the system feed on each other over time. To counter this, HiCore builds three families of hypergraphs (item-, entity-, and word-oriented), each from three channels (group, joint, purchase), to capture multi-level user interests. A self-supervised InfoNCE objective fuses these interests, and the resulting representation is injected into both the recommendation and conversation modules. The authors claim that HiCore outperforms all baselines on four benchmarks and that its higher coverage and lower average popularity demonstrate real mitigation of the Matthew effect in this dynamic setting.

Core claim

On its own terms, the paper's discovery is that constructing a diverse set of hypergraphs—over items, knowledge-graph entities, and conversational words, with each split into group, joint, and purchase channels—lets a conversational recommender learn broader user interests than single-channel hypergraphs or pairwise graphs. These interests, learned with a self-supervised contrastive objective, then feed both item prediction and response generation. The empirical claim is that this multi-interest signal improves recommendation accuracy and dialogue diversity simultaneously, and that the increased coverage and lower popularity of recommended items show the Matthew effect being mitigated in the

What carries the argument

The load-bearing component is the triple-channel multi-hypergraph: item-, entity-, and word-oriented hypergraphs built from network motifs and grouped into group, joint, and purchase channels. Hypergraph convolution propagates user embeddings over each channel, attention combines the channels into item-, entity-, and word-level interest representations, and an InfoNCE self-supervised loss forces these representations to be robust. A gated fusion produces a single multi-interest embedding that is inserted into both the recommendation classifier and the Transformer-based conversation generator.

Load-bearing premise

The claim that HiCore mitigates the dynamic Matthew effect rests on static offline diversity metrics (coverage, average popularity, long-tail ratio) serving as proxies for a dynamic user-system feedback loop; if these proxies do not capture the loop, the central conclusion is unsupported.

What would settle it

Run a user-agent simulation (or an online A/B test) where recommendations from HiCore and a baseline are repeatedly fed back into the user model, and track the popularity concentration of recommended items over turns; the Matthew-effect claim would be falsified if HiCore's recommendations become increasingly popular over time, or if its lower Long Tail Ratio is shown to be an artifact of recommending fewer long-tail items.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If HiCore's claim holds, multi-channel hypergraph interest modeling is a viable route to simultaneously improve accuracy and diversification in conversational recommendation.
  • The framework's success would suggest that going beyond pairwise item-item edges to higher-order motifs (group, joint, purchase) helps uncover user interests that single-channel hypergraphs miss.
  • The improved dialogue diversity (higher Distinct-n) would imply that richer interest representations can make conversational agents produce more varied responses while still recommending accurately.
  • The reported higher coverage and lower average popularity indicate that the method spreads recommendations across a broader item space, which the paper interprets as counteracting the popularity feedback loop.
  • Ablations showing that removing any of the nine hypergraphs hurts performance imply that each channel contributes non-redundant signal to the final interest representation.

Where Pith is reading between the lines

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

  • The Matthew-effect claim is supported only by static offline diversity metrics; a true test would require an online or simulated loop where the system's recommendations influence which items users engage with next, then measuring whether popularity concentration grows over time.
  • The paper interprets a lower Long Tail Ratio as evidence of mitigating the Matthew effect, but in standard usage a lower ratio typically means fewer long-tail items are recommended—so this metric may actually point in the opposite direction unless the paper uses a nonstandard definition.
  • The multi-interest representation could be transferred to other interactive settings (e.g., sequential recommendation, conversational search) where interest confinement is the suspected driver of bias.
  • A direct extension would be to add a causal or counterfactual correction for popularity bias on top of the hypergraph interests, which might make the diversity gains robust to distribution shift when deployed online.

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

4 major / 6 minor

Summary. The manuscript proposes HiCore, a multi-hypergraph multi-interest self-supervised learning framework for conversational recommendation (CRS). It builds item-, entity-, and word-oriented hypergraphs from group, joint, and purchase channels, learns multi-level user interests via hypergraph convolution and an InfoNCE-style contrastive objective, and feeds the resulting interest representation into both recommendation and conversation modules. The paper claims state-of-the-art results on REDIAL, TG-REDIAL, OpenDialKG, and DuRecDial, and claims to mitigate the Matthew effect in the dynamic user-system feedback loop. Experiments include recommendation metrics (Recall/MRR/NDCG), dialogue distinctness, coverage, average popularity, long-tail ratio, hyperparameter analysis, and ablations.

Significance. If the results and the Matthew-effect framing were sound, HiCore would be a relevant contribution to CRS and to diversity-oriented recommendation: it introduces a multi-channel hypergraph construction, a multi-interest contrastive objective, and a joint recommendation/conversation architecture, with code released. However, the central empirical claim is internally inconsistent: Table 1 reports mathematically impossible MRR > Recall values, and the same table contradicts the claim of consistent improvement over baselines. The Matthew-effect conclusion is based on static proxy metrics whose direction is not clearly justified, and the equal-title prior EMNLP 2024 publication in the reference list removes the novelty claim. Hence the significance is currently not established by this manuscript.

major comments (4)
  1. [§4.2, Table 1] The TG-REDIAL row for HiCore reports M@10=0.0880 > R@10=0.0270 and M@50=0.1074 > R@50=0.0769. With one ground-truth item per test dialogue, MRR@K is the mean of 1/rank when rank≤K and 0 otherwise, so MRR@K ≤ Recall@K for every test case. The reported inequality is therefore impossible under the stated protocol. The same row also has R@10 and R@50 below MHIM (0.0270 vs 0.0300; 0.0769 vs 0.0783), contradicting the claim in §4.2 that HiCore 'consistently outperforms all the comparison baselines' and the asterisk indicating p<0.05 over all baselines. Table 2 similarly shows HiCore below HyCoRec on OpenDialKG R@1 (0.2628 vs 0.2742). This internal inconsistency invalidates the empirical cornerstone of the SOTA claim.
  2. [§4.4, Table 4] The Matthew-effect claim rests on static offline metrics: average popularity A@K, long-tail ratio L@K, and coverage. No dynamic simulation or online interaction is presented, so the paper has no direct evidence about the 'dynamic user-system feedback loop' stated in the abstract. Moreover, the interpretation of L@K is questionable: if L@K is the fraction of recommended long-tail items, lower HiCore values (e.g., 0.1906 vs MHIM 0.1919 on OpenDialKG L@5) mean fewer long-tail items are recommended, the opposite of 'successfully addresses the long tail distribution.' Since the metric is not formally defined, the conclusion that lower L@K indicates Matthew-effect mitigation is unsupported and appears inverted.
  3. [References / §1] The reference list contains [Zheng et al., 2024f], an EMNLP 2024 paper with the same title, same author group, and the same proposed framework HiCore. The current manuscript does not state that it is an extension, republication, or overlapping work, and §1 claims 'this is the first work' to build such hypergraphs. The central novelty and contribution therefore appear to have been published previously. This is a load-bearing disclosure issue and needs to be resolved (e.g., by clear deltas over the prior version, or by withdrawal if the submission is a duplicate).
  4. [§3.1.2, Eq. (9)] The purchase-channel adjacency is defined as A_p = A_M10 − (A_M8 + A_M9). Subtracting motif adjacency matrices is not a standard hypergraph construction; unless the joint motifs are strict subgraphs of the purchase motif, entries can be negative or zero in ways that change the degree-normalized propagation in Eq. (8). No proof or motivation is given that this subtraction yields a meaningful hypergraph, and the downstream contrastive learning relies on this matrix. Please justify or replace with a nonnegative construction.
minor comments (6)
  1. [Abstract] Typo: 'Matthew effec' should be 'Matthew effect'.
  2. [Figure 1] The caption and diagram use 'Interested-Boosted CRS'; should be 'Interest-Boosted CRS'.
  3. [Eq. (11)] The objective has unmatched parentheses and inconsistent log placement in the two contrastive terms; the definitions of z_h^u, \hat z_h^u, and k_h should also be made more precise.
  4. [§4.3] The sentence beginning 'It also gains 446.51%...' is attributed to the REDIAL dataset, but the numbers appear to correspond to TG-REDIAL; the dataset name is likely repeated in error.
  5. [Eq. (14)] p2 and p3 are both conditioned on Prec; presumably one condition is a different context. Please clarify.
  6. [References] Deng et al. 2021a and 2021b are duplicate entries for the same paper; please consolidate.

Circularity Check

0 steps flagged

No significant circularity: HiCore's SOTA claim is an empirical benchmark result, not an analytic consequence of its own definitions.

full rationale

I walked the derivation chain. The recommendation score is computed as P_rec = X_m × V_cand, where X_m is produced by hypergraph convolution and attention over the constructed hypergraphs, and the SSL loss in Eq. (11) is a standard contrastive objective: positive pairs are different views of the same user data, with the text saying X_m and Z_h are 'ground truths for each other.' That is a description of contrastive learning, not a validation loop that makes the reported benchmark numbers true by construction. No parameter is fitted to the Matthew-effect metrics (Coverage, A@K, L@K) and then reported as a prediction; those metrics are measured on outputs of a model trained with cross-entropy and InfoNCE losses. The heavy self-citation, including the authors' identical EMNLP 2024 entry and HyCoRec, is unusual but not load-bearing: the central comparison is against externally published baselines, and the framework's components are defined from the data rather than from those citations. The impossible-looking Table 1 cell (M@10 > R@10 on TG-REDIAL) and the questionable use of offline long-tail metrics as evidence for a dynamic feedback-loop effect are correctness and validity concerns, not circularity. Under the required standard of exhibiting a specific reduction to inputs by construction, no such circular step is present.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The framework introduces no new physical or formal entities; its load-bearing assumptions are domain claims about motif-based hypergraphs and about offline metrics proxying dynamic feedback-loop effects. The free parameters are all standard model hyperparameters, but none are reported with final values, which weakens reproducibility.

free parameters (5)
  • Embedding dimension d
    Tuned via Fig. 4; final value not reported. Affects all embeddings and the final recommendation score Prec.
  • Self-supervised/conversation weight beta
    Hyperparameter in Eq. (13) and the self-supervised objective; discussed in Fig. 4 but no final value given.
  • Hypergraph convolution layers N = 2 (best in Fig. 4)
    Sec. 4.5 states a two-layer hypergraph convolution is sufficient; exact layer count is a tuned choice.
  • Hyperedge threshold P
    Mentioned in RQ4 and Fig. 4 but never defined or reported; controls the number of hyperedges.
  • k-hop neighbor size for entity/word hypergraphs
    k is used to construct entity- and word-oriented hyperedges in Sec. 3.1.1, but no value is stated.
axioms (5)
  • standard math Hypergraph convolution propagation in Eq. (8) is a valid encoder for user interests.
    Taken from Yu et al. (2021) without proof and used as the core learning operator.
  • domain assumption Triangle motifs defined on social relations encode group, joint, and purchase channels relevant to recommendation interests.
    Sec. 3.1.1 assumes these motif structures capture user preferences and social dynamics.
  • domain assumption Multi-level interests learned from static offline logs transfer to dynamic user-system feedback loops.
    The paper motivates CRS dynamics in Sec. 1 but evaluates only static offline datasets.
  • ad hoc to paper Lower Average Popularity and lower Long Tail Ratio indicate successful Matthew-effect mitigation.
    Sec. 4.4 and Table 4 treat the lowest L@K as success; this is not standard and is internally questionable.
  • ad hoc to paper Subtraction in the purchase-channel adjacency (A_p = A_M10 - (A_M8 + A_M9), Eq. 9) yields a meaningful hypergraph.
    No justification is given for subtracting adjacency matrices; negative entries are possible.

pith-pipeline@v1.3.0-alltime-deepseek · 16265 in / 14650 out tokens · 139988 ms · 2026-08-01T14:50:24.682794+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation." pith.science (2026). https://pith.science/paper/OBQRUP5Y

@misc{pith2026260718609,
  author       = {Pith},
  title        = {Pith review of: Mitigating Matthew Effect: Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OBQRUP5Y}},
  note         = {Machine review of arXiv:2607.18609}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The Matthew effect is a big challenge in Recommender Systems (RSs), where popular items tend to receive increasing attention, while less popular ones are often overlooked, perpetuating existing disparities. Although many existing methods attempt to mitigate Matthew effect in the static or quasi-static recommendation scenarios, such issue will be more pronounced as users engage with the system over time. To this end, we propose a novel framework, Multi-Hypergraph Boosted Multi-Interest Self-Supervised Learning for Conversational Recommendation (HiCore), aiming to address Matthew effect in the Conversational Recommender System (CRS) involving the dynamic user-system feedback loop. It devotes to learn multi-level user interests by building a set of hypergraphs (i.e., item-, entity-, word-oriented multiple-channel hypergraphs) to alleviate the Matthew effec. Extensive experiments on four CRS-based datasets showcase that HiCore attains a new state-of-the-art performance, underscoring its superiority in mitigating the Matthew effect effectively. Our code is available at https://github.com/zysensmile/HiCore.

Figures

Figures reproduced from arXiv: 2607.18609 by Guohua Wang, Kwok-Yan Lam, Liang Lin, Ruilin Xu, Yongsen Zheng.

Figure 1
Figure 1. Figure 1: Overview of our HiCore framework. It consists of Multi-Hypergraph Boosted Multi-Interest Self [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Triangle motifs used in our proposed HiCore. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coverage results of C@k metric. constructing a diverse set of hypergraphs, encom￾passing item-oriented, entity-oriented, and word￾oriented triple-channel hypergraphs. These struc￾tures effectively capture intricate local patterns through motif analysis, enabling the exploration of high-order user behaviors. This proves invalu￾able in generating informative and high-quality re￾sponse utterances. (b) HiCore … view at source ↗
Figure 4
Figure 4. Figure 4: Impact of different hyperparameteres. strates the lowest values for Average Popularity and Long Tail Ratio. This evidence suggests that our method effectively mitigates the adverse effects of item popularity on recommendation outcomes and successfully addresses the long tail distribution of items. These results validate the effectiveness of our proposed approach in combating the Matthew effect in the CRS a… 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

116 extracted references · 4 canonical work pages

  1. [1]

    Empirical Methods in Natural Language Processing (Demonstrations) , pages=

    Openke: An open toolkit for knowledge embedding , author=. Empirical Methods in Natural Language Processing (Demonstrations) , pages=

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Translating embeddings for modeling multi-relational data , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    IEEE Transactions on Information Theory , volume=

    On the role of dynamic programming in statistical communication theory , author=. IEEE Transactions on Information Theory , volume=. 1957 , publisher=

    1957

  4. [4]

    Conference on Research and Development in Information Retrieval , pages=

    Interactive recommender system via knowledge graph-enhanced reinforcement learning , author=. Conference on Research and Development in Information Retrieval , pages=

  5. [5]

    Association for the Advancement of Artificial Intelligence , volume=

    Deep reinforcement learning with double q-learning , author=. Association for the Advancement of Artificial Intelligence , volume=

  6. [6]

    The Semantic Web: 15th International Conference , pages=

    Modeling relational data with graph convolutional networks , author=. The Semantic Web: 15th International Conference , pages=. 2018 , organization=

    2018

  7. [7]

    Advances in Neural Information Processing Systems , volume=

    Iterative deep graph learning for graph neural networks: Better and robust node embeddings , author=. Advances in Neural Information Processing Systems , volume=

  8. [8]

    arXiv preprint arXiv:1908.04942 , year=

    Reinforcement learning based graph-to-sequence model for natural question generation , author=. arXiv preprint arXiv:1908.04942 , year=

    Pith/arXiv arXiv 1908

  9. [9]

    arXiv preprint arXiv:1609.02907 , year=

    Semi-supervised classification with graph convolutional networks , author=. arXiv preprint arXiv:1609.02907 , year=

    Pith/arXiv arXiv

  10. [10]

    The Web Conference , pages=

    Conversational contextual bandit: Algorithm and application , author=. The Web Conference , pages=

  11. [11]

    ACM Transactions on Information Systems , volume=

    Seamlessly unifying attributes and items: Conversational recommendation for cold-start users , author=. ACM Transactions on Information Systems , volume=. 2021 , publisher=

    2021

  12. [12]

    arXiv preprint arXiv:2010.04125 , year=

    Towards topic-guided conversational recommender system , author=. arXiv preprint arXiv:2010.04125 , year=

    Pith/arXiv arXiv 2010

  13. [13]

    International Conference on Knowledge Discovery and Data Mining , pages=

    Improving conversational recommender systems via knowledge graph based semantic fusion , author=. International Conference on Knowledge Discovery and Data Mining , pages=

  14. [14]

    arXiv preprint arXiv:2005.03954 , year=

    Towards conversational recommendation over multi-type dialogs , author=. arXiv preprint arXiv:2005.03954 , year=

    Pith/arXiv arXiv 2005

  15. [15]

    Conference on Research and Development in Information Retrieval , pages=

    Towards question-based recommender systems , author=. Conference on Research and Development in Information Retrieval , pages=

  16. [16]

    Conference on Research and Development in Information Retrieval , pages=

    Conversational recommendation: Formulation, methods, and evaluation , author=. Conference on Research and Development in Information Retrieval , pages=

  17. [17]

    Conference on Information and Knowledge Management , pages=

    Towards conversational search and recommendation: System ask, user respond , author=. Conference on Information and Knowledge Management , pages=

  18. [18]

    International Conference on Knowledge Discovery and Data Mining , pages=

    A visual dialog augmented interactive recommender system , author=. International Conference on Knowledge Discovery and Data Mining , pages=

  19. [19]

    International Conference on Multimedia , pages=

    Knowledge-aware multimodal dialogue systems , author=. International Conference on Multimedia , pages=

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Towards deep conversational recommendations , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    International Conference on Knowledge Discovery and Data Mining , pages=

    Towards conversational recommender systems , author=. International Conference on Knowledge Discovery and Data Mining , pages=

  22. [22]

    International Conference on Knowledge Discovery and Data Mining , pages=

    Q&R: A two-stage approach toward interactive recommendation , author=. International Conference on Knowledge Discovery and Data Mining , pages=

  23. [23]

    arXiv preprint arXiv:1908.05391 , year=

    Towards knowledge-based recommender dialog system , author=. arXiv preprint arXiv:1908.05391 , year=

    Pith/arXiv arXiv 1908

  24. [24]

    The Web Conference , pages=

    Multiple Choice Questions based Multi-Interest Policy Learning for Conversational Recommendation , author=. The Web Conference , pages=

  25. [25]

    Conference on Research and Development in Information Retrieval , pages=

    Learning to Infer User Implicit Preference in Conversational Recommendation , author=. Conference on Research and Development in Information Retrieval , pages=

  26. [26]

    Conference on Research and Development in Information Retrieval , pages=

    Conversational Recommendation via Hierarchical Information Modeling , author=. Conference on Research and Development in Information Retrieval , pages=

  27. [27]

    Conference on Research and Development in Information Retrieval , pages=

    Unified conversational recommendation policy learning via graph-based reinforcement learning , author=. Conference on Research and Development in Information Retrieval , pages=

  28. [28]

    Conference on Research and Development in Information Retrieval , pages=

    Learning to ask appropriate questions in conversational recommendation , author=. Conference on Research and Development in Information Retrieval , pages=

  29. [29]

    Web Search and Data Mining , pages=

    Adapting user preference to online feedback in conversational recommendation , author=. Web Search and Data Mining , pages=

  30. [30]

    International Conference on Knowledge Discovery and Data Mining , pages=

    Interactive path reasoning on graph for conversational recommendation , author=. International Conference on Knowledge Discovery and Data Mining , pages=

  31. [31]

    Web Search and Data Mining , pages=

    Estimation-action-reflection: Towards deep interaction between conversational and recommender systems , author=. Web Search and Data Mining , pages=

  32. [32]

    Conversational recommender system , author=

  33. [33]

    2021 , publisher=

    Advances and challenges in conversational recommender systems: A survey , author=. 2021 , publisher=

    2021

  34. [34]

    Hongshen Chen and Xiaorui Liu and Dawei Yin and Jiliang Tang , title =

  35. [35]

    Findings of the Association for Computational Linguistics , pages =

    Yu Lu and Junwei Bao and Yan Song and Zichen Ma and Shuguang Cui and Youzheng Wu and Xiaodong He , title =. Findings of the Association for Computational Linguistics , pages =

  36. [36]

    Web Search and Data Mining , pages =

    Yuanhang Zhou and Kun Zhou and Wayne Xin Zhao and Cheng Wang and Peng Jiang and He Hu , title =. Web Search and Data Mining , pages =

  37. [37]

    International Conference on Knowledge Discovery and Data Mining , pages =

    Konstantina Christakopoulou and Filip Radlinski and Katja Hofmann , title =. International Conference on Knowledge Discovery and Data Mining , pages =

  38. [38]

    Conference on Research

    Yueming Sun and Yi Zhang , title =. Conference on Research

  39. [39]

    Leveraging Historical Interaction Data for Improving Conversational Recommender System , booktitle =

    Kun Zhou and Wayne Xin Zhao and Hui Wang and Sirui Wang and Fuzheng Zhang and Zhongyuan Wang and Ji. Leveraging Historical Interaction Data for Improving Conversational Recommender System , booktitle =

  40. [40]

    The Web Conference , pages =

    Yiming Zhang and Lingfei Wu and Qi Shen and Yitong Pang and Zhihua Wei and Fangli Xu and Bo Long and Jian Pei , title =. The Web Conference , pages =

  41. [41]

    Empirical Methods in Natural Language Processing , pages =

    Shirley Anugrah Hayati and Dongyeop Kang and Qingxiaoyang Zhu and Weiyan Shi and Zhou Yu , title =. Empirical Methods in Natural Language Processing , pages =

  42. [42]

    Advances in Neural Information Processing Systems , pages =

    Raymond Li and Samira Ebrahimi Kahou and Hannes Schulz and Vincent Michalski and Laurent Charlin and Chris Pal , title =. Advances in Neural Information Processing Systems , pages =

  43. [43]

    Towards Topic-Guided Conversational Recommender System , booktitle =

    Kun Zhou and Yuanhang Zhou and Wayne Xin Zhao and Xiaoke Wang and Ji. Towards Topic-Guided Conversational Recommender System , booktitle =

  44. [44]

    Conference on Research and Development in Information Retrieval , pages =

    Zhihui Xie and Tong Yu and Canzhe Zhao and Shuai Li , title =. Conference on Research and Development in Information Retrieval , pages =

  45. [45]

    Association for Computational Linguistics , pages =

    Xiexiong Lin and Weiyu Jian and Jianshan He and Taifeng Wang and Wei Chu , title =. Association for Computational Linguistics , pages =

  46. [46]

    Web Search and Data Mining , pages =

    Xiaoyu Zhang and Xin Xin and Dongdong Li and Wenxuan Liu and Pengjie Ren and Zhumin Chen and Jun Ma and Zhaochun Ren , title =. Web Search and Data Mining , pages =

  47. [47]

    Conference on Research and Development in Information Retrieval , pages =

    Shuokai Li and Ruobing Xie and Yongchun Zhu and Xiang Ao and Fuzhen Zhuang and Qing He , title =. Conference on Research and Development in Information Retrieval , pages =

  48. [48]

    Yingxu Wang and Xiaoru Chen and Jinyuan Fang and Zaiqiao Meng and Shangsong Liang , title =

  49. [49]

    Association for the Advancement of Artificial Intelligence , pages =

    Filip Radlinski and Craig Boutilier and Deepak Ramachandran and Ivan Vendrov , title =. Association for the Advancement of Artificial Intelligence , pages =

  50. [50]

    DBpedia:

    S. DBpedia:. International Semantic Web Conference/Asian Semantic Web Conference , volume =

  51. [51]

    Conference on Research and Development in Information Retrieval , pages =

    Yang Deng and Yaliang Li and Fei Sun and Bolin Ding and Wai Lam , title =. Conference on Research and Development in Information Retrieval , pages =

  52. [52]

    Pointing the Unknown Words , booktitle =

  53. [53]

    Gomez and Lukasz Kaiser and Illia Polosukhin , title =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , title =. Advances in Neural Information Processing Systems , pages =

  54. [54]

    CoRR , volume =

    Tong Zhang and Yong Liu and Peixiang Zhong and Chen Zhang and Hao Wang and Chunyan Miao , title =. CoRR , volume =

  55. [55]

    Conference on Recommender Systems , pages =

    Harald Steck , title =. Conference on Recommender Systems , pages =

  56. [56]

    ACM Transactions on Information Systems , year =

    Chongming Gao and Wenqiang Lei and Jiawei Chen and Shiqi Wang and Xiangnan He and Shijun Li and Biao Li and Yuan Zhang and Peng Jiang , title =. ACM Transactions on Information Systems , year =

  57. [57]

    ACM on Human-Computer Interaction , volume =

    Eslam Hussein and Prerna Juneja and Tanushree Mitra , title =. ACM on Human-Computer Interaction , volume =

  58. [58]

    Shapiro and Mustafa Bilgic , title =

    Ping Liu and Karthik Shivaram and Aron Culotta and Matthew A. Shapiro and Mustafa Bilgic , title =. The Web Conference , pages =

  59. [59]

    Nguyen and Pik

    Tien T. Nguyen and Pik. Exploring the filter bubble: the effect of using recommender systems on content diversity , booktitle =

  60. [60]

    Auditing radicalization pathways on YouTube , booktitle =

    Manoel Horta Ribeiro and Raphael Ottoni and Robert West and Virg. Auditing radicalization pathways on YouTube , booktitle =

  61. [61]

    Conference on User Modeling, Adaptation, and Personalization , pages =

    Larissa Spinelli and Mark Crovella , title =. Conference on User Modeling, Adaptation, and Personalization , pages =

  62. [62]

    An Audit of Misinformation Filter Bubbles on YouTube: Bubble Bursting and Recent Behavior Changes , booktitle =

    Mat. An Audit of Misinformation Filter Bubbles on YouTube: Bubble Bursting and Recent Behavior Changes , booktitle =

  63. [63]

    Association for the Advancement of Artificial Intelligence , pages =

    Libo Qin and Zhouyang Li and Qiying Yu and Lehan Wang and Wanxiang Che , title =. Association for the Advancement of Artificial Intelligence , pages =

  64. [64]

    Association for the Advancement of Artificial Intelligence , pages =

    Shimin Li and Qinyuan Cheng and Linyang Li and Xipeng Qiu , title =. Association for the Advancement of Artificial Intelligence , pages =

  65. [65]

    Association for the Advancement of Artificial Intelligence , pages =

    Kshitij Mishra and Priyanshu Priya and Asif Ekbal , title =. Association for the Advancement of Artificial Intelligence , pages =

  66. [66]

    Alleviating the Long-Tail Problem in Conversational Recommender Systems , booktitle =

    Zhipeng Zhao and Kun Zhou and Xiaolei Wang and Wayne Xin Zhao and Fan Pan and Zhao Cao and Ji. Alleviating the Long-Tail Problem in Conversational Recommender Systems , booktitle =

  67. [67]

    Epure and Romain Hennequin , title =

    Elena V. Epure and Romain Hennequin , title =. European Chapter of the Association for Computational Linguistics , pages =

  68. [68]

    Conversational Agents for Mental Health and Well-being: Discovering Design Recommendations Using Text Mining , booktitle =

    Abdullah Wahbeh and Mohammad Al. Conversational Agents for Mental Health and Well-being: Discovering Design Recommendations Using Text Mining , booktitle =

  69. [69]

    Conference on Research and Development in Information Retrieval , pages =

    Yuanxing Liu and Weinan Zhang and Baohua Dong and Yan Fan and Hang Wang and Fan Feng and Yifan Chen and Ziyu Zhuang and Hengbin Cui and Yongbin Li and Wanxiang Che , title =. Conference on Research and Development in Information Retrieval , pages =

  70. [70]

    Association for the Advancement of Artificial Intelligence , pages =

    Robyn Speer and Joshua Chin and Catherine Havasi , title =. Association for the Advancement of Artificial Intelligence , pages =

  71. [71]

    Empirical Methods in Natural Language Processing (Demonstrations) , pages =

    Yoon Kim , title =. Empirical Methods in Natural Language Processing (Demonstrations) , pages =

  72. [72]

    User-controllable Recommendation Against Filter Bubbles , booktitle =

    Wenjie Wang and Fuli Feng and Liqiang Nie and Tat. User-controllable Recommendation Against Filter Bubbles , booktitle =

  73. [73]

    The Web Conference , pages =

    Zhenyang Li and Yancheng Dong and Chen Gao and Yizhou Zhao and Dong Li and Jianye Hao and Kai Zhang and Yong Li and Zhi Wang , title =. The Web Conference , pages =

  74. [74]

    Examining the Matthew Effect on YouTube Recommendation System , year=

    Liu, Ying Chieh and Huang, Min Qi , booktitle=. Examining the Matthew Effect on YouTube Recommendation System , year=

  75. [75]

    Conference on Research and Development in Information Retrieval , pages =

    Chongming Gao and Kexin Huang and Jiawei Chen and Yuan Zhang and Biao Li and Peng Jiang and Shiqi Wang and Zhong Zhang and Xiangnan He , title =. Conference on Research and Development in Information Retrieval , pages =

  76. [76]

    The Web Conference , pages =

    Ashton Anderson and Lucas Maystre and Ian Anderson and Rishabh Mehrotra and Mounia Lalmas , title =. The Web Conference , pages =

  77. [77]

    Conference on Web Search and Data Mining , pages =

    Christian Hansen and Rishabh Mehrotra and Casper Hansen and Brian Brost and Lucas Maystre and Mounia Lalmas , title =. Conference on Web Search and Data Mining , pages =

  78. [78]

    Conference on Research and Development in Information Retrieval , pages =

    Yile Liang and Tieyun Qian and Qing Li and Hongzhi Yin , title =. Conference on Research and Development in Information Retrieval , pages =

  79. [79]

    The Web Conference , pages =

    Yu Zheng and Chen Gao and Liang Chen and Depeng Jin and Yong Li , title =. The Web Conference , pages =

  80. [80]

    The Web Conference , pages =

    Yu Zheng and Chen Gao and Xiang Li and Xiangnan He and Yong Li and Depeng Jin , title =. The Web Conference , pages =

Showing first 80 references.