Counterfactual Multi-player Bandits for Explainable Recommendation Diversification

Bowei He; Chen Ma; Haolun Wu; Xiaokun Zhang; Yansen Zhang; Zexu Sun

REVIEW 4 major objections 6 minor 1 cited by

Counterfactual Multi-player Bandits for Explainable Recommendation Diversification

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Bandits find the minimal feature tweaks that diversify recommendations.

desk verdict Novel combination, but test-set ambiguity and a false independence assumption undercut the main claims. read the letter →

arxiv 2505.21165 v2 pith:YRDPMCIK submitted 2025-05-27 cs.IR

Yansen Zhang , Bowei He , Xiaokun Zhang , Haolun Wu , Zexu Sun , Chen Ma This is my paper · ORCID

classification cs.IR

keywords diversifiedrecommendationcounterfactualframeworkmulti-armedbanditsnon-differentiablediversitymetricsaccuracy-diversitytrade-offfeature-levelexplanationfilterbubble

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

Recommenders trained only for accuracy can trap users in filter bubbles, but existing diversity fixes are hard to explain and often optimize only differentiable metrics. This paper proposes a counterfactual multi-player bandit (CMB) method that searches over small perturbations to item feature vectors, seeking the minimal changes that raise a chosen diversity score. The same search works for differentiable and non-differentiable diversity metrics because it uses reward feedback instead of gradients. The paper claims that, on MovieLens 1M, MovieLens 10M, and Amazon CDs, CMB improves diversity with much smaller accuracy losses than MMR or DPP, and that the features receiving larger perturbations are the ones that drive diversification.

What carries the argument

The carrier of the argument is the perturbation matrix $\Delta$ together with a multi-player bandit search over its entries. Each item-feature entry is one player; the arms are candidate perturbation values sampled evenly from $[-A,A]$, and an $\varepsilon$-greedy rule selects arms using an incremental-average estimate of each arm's reward. The reward is the counterfactual objective of Eq. 9 (or the accuracy-diversity version of Eq. 14), so the method is gradient-free and metric-agnostic. An $\ell_1$ penalty on $\Delta$ encourages minimal interventions, making the final nonzero entries a sparse, interpretable explanation of what drives diversity.

What would settle it

On a small item set, run CMB for a fixed diversity metric and compare its final $\Delta$ with the $\Delta$ found by exhaustive enumeration of all perturbation combinations that satisfy the same $\ell_1$ budget; if the bandit's $\Delta$ is far from the true maximizer, or if adding an explicit coupling term to the reward estimator changes the selected features, the independence assumption is doing the work.

Watch

Extended reading notes

Core claim

CMB treats the item latent feature matrix $Q$ as an intervention surface. It learns a perturbation matrix $\Delta$ such that $Q+\Delta$, fed through a fixed base model $g$, makes the top-$K$ lists more diverse while keeping accuracy close to the original. Each entry $\Delta_{i,j}$ is a 'player' choosing an arm from a discrete set of perturbation values, and the shared reward is the counterfactual objective: either a single diversity metric or a weighted combination of accuracy and diversity. Because the objective is only evaluated as a reward, the method optimizes metrics like $\alpha$-nDCG, subtopic coverage, prediction coverage, and ILAD directly, without gradients. After convergence, the magnitude of $\Delta$ provides the explanation: features with larger absolute perturbations are the ones whose change yields more diverse recommendations. Experiments on three datasets are reported as showing that CMB achieves a better accuracy-diversity trade-off than MMR and DPP, and that erasing the identified top features removes most of the diversity gain.

Load-bearing premise

The bandit update assumes every item-feature perturbation can be scored independently of the others, when in fact the reward is one global diversity score that shifts whenever any entry of $\Delta$ changes; if that assumption fails, the per-arm estimates are biased.

Editorial extensions

If this is right

Any diversity metric that can be computed as a number on a recommendation list can be optimized directly, including non-differentiable ones such as prediction coverage and subtopic coverage.
The method is model-agnostic: the same search works with BPRMF and LightGCN, and the paper's case study extends it to raw item features.
The magnitude of $\Delta$ gives an actionable feature-level explanation: system designers can see which dimensions of item representations push diversity upward.
Combining accuracy and diversity in the objective yields smaller accuracy loss than optimizing diversity alone, while still improving diversity scores.
Because the search only perturbs the input to a fixed model, it can be applied after training without retraining the recommender.

Reading between the lines

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

If the identified features transfer across list lengths or datasets, practitioners could pre-emptively regularize or steer those latent dimensions instead of re-ranking after the fact.
The bandit formulation could be extended to per-user perturbation budgets, testing whether 'diversity drivers' differ across user segments.
An ablation comparing CMB with a fully coupled joint search, or with a contextual bandit that conditions on other players' arms, would isolate how much of the reported trade-off depends on the independence assumption.
The same counterfactual search could be used to explain other list-level properties, such as popularity bias or serendipity, by swapping the reward function.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Colleague,

The paper proposes CMB, a counterfactual perturbation of item embeddings jointly optimized by a collection of epsilon-greedy bandits, to optimize diversity metrics directly and to identify feature-level explanations. The combination is new and the framing of each perturbation entry as a player is unusual. The tool is genuinely model-agnostic and metric-agnostic, and the authors back it with experiments on three datasets, a gradient baseline, a random baseline, and released code.

Where it struggles: the evaluation has a potentially fatal ambiguity. The paper never states which split is used when learning the perturbation Δ in Eq. 9. Because the objective is the same diversity metric used for evaluation, if the perturbation is learned on the test split, the reported diversity gains are fitted rather than predicted. This is the first thing a referee would need resolved.

Second, the bandit rests on an explicit independence assumption (Section 3.3) that is false. The reward in Eq. 9 depends on the full Δ through the top-K rankings, so the reward credited to a single player's arm also reflects every other player's choices. The incremental-average update in Eq. 13 is therefore not estimating a stationary arm value, and epsilon-greedy selection is not optimizing Eq. 9 in any well-defined sense. The text acknowledges the assumption but does not justify or test it.

Third, the random-arm baseline captures most of the diversity gain: on ML10M, random gets ILAD@10 of 0.2584 versus 0.2652 for the learned perturbation; on CDs, 0.3081 versus 0.3102. That suggests the learned Δ's structure is not clearly load-bearing, and it undercuts the feature-importance explanation built on it. The erasure test in Section 4.3 is also partly circular: erasing the largest components of a perturbation trained to increase diversity will naturally reduce diversity.

The paper is not sloppy; it is clearly written, it reports the baseline that hurts its own claims, and it releases code. But the central empirical claim and the explanation claim are not adequately supported as written. The split ambiguity alone could send it to rejection; the bandit soundness issue makes it worse.

I would not desk-reject it outright, but I would not accept it in its current form. It is best read by people working on diversity optimization and counterfactual explanation, with a skeptical eye on the experimental protocol. I would not cite it in its current form.

Recommendation: reject, with the issues stated; a resubmission after experiments on a proper split and a sounder treatment of the bandit could change the picture.

Referee Report

4 major / 6 minor

Summary. The paper proposes CMB, a counterfactual framework that learns a perturbation matrix Δ on the item latent feature matrix Q of a pre-trained recommender. Each entry Δ_{i,j} is treated as a player in a multi-player bandit, and the players jointly select arm values from a finite grid to maximize a reward given by a diversity metric (Eq. 9) or a weighted accuracy-diversity trade-off (Eq. 14). After optimization, the magnitude of Δ is interpreted as feature-level importance and is validated by an erasure test. Experiments on ML1M, ML10M, and CDs with BPRMF and LightGCN compare CMB with MMR and DPP on Recall, NDCG, α-nDCG, SC, PC, and ILAD.

Significance. If the proposed mechanism were valid, the metric-agnostic, model-agnostic counterfactual approach would be a useful addition to the diversification literature, and the paper deserves credit for releasing code, including a random-arm baseline and a gradient variant, and evaluating against several diversity metrics. However, the central optimization procedure in Section 3.3 is not a valid bandit algorithm because the joint reward couples all players, the random baseline in Table 5 achieves similar diversity gains, and the explanation validity test in Section 4.3 is circular; the empirical claims are therefore not established.

major comments (4)

[Section 3.3, Eqs. (9) and (13)] The multi-player bandit model is misspecified. The reward r_t equals eΨ, which is computed from the ranking lists produced by the full perturbation matrix Δ assembled from all d×|V| players. Eq. (13) nevertheless updates each player's arm value V_{Δ_p} by the incremental average of this shared reward, which is only valid if each arm has a stationary value independent of other players' choices. The paper explicitly assumes 'different players are independent of each other', but this assumption is false: changing Δ_{i,j} for one item feature alters the top-K lists and therefore eΨ for every player. The reward is non-stationary and coupled across players, so the per-arm estimates are biased and the ε-greedy selection does not optimize Eq. (9) in any well-defined sense. Consequently, the learned Δ cannot be claimed to be a minimal intervention or a maximizer of the counterfactual objective, which undermines the optimization mechanism behind RQ1.
[Table 5 and Section 4.2] The random-arm baseline CMBBPRMF-Random achieves nearly the same diversity as the learned single-objective variants on ML10M (ILAD@10: 0.2584 vs 0.2652 for CMBBPRMF-ILAD; PC@10: 0.2625 vs 0.2714), and on CDs it even beats several learned variants on PC@10 and ILAD@10. This observation, together with the invalid bandit update in Eq. (13), indicates that the diversity gains over BPRMF may be attributable to adding bounded random perturbations rather than to the bandit search. The statement that 'our bandit method is effective for optimizing different diversity metrics' is therefore not supported by the evidence presented.
[Sections 4.1–4.2, Eq. (14)] The manuscript does not state on which data split the Accuracy(eR_g) term in the counterfactual reward is computed. The evaluation protocol is described for test labels, and if the reward uses test ground truth, then Δ is learned by maximizing a function of test accuracy and test diversity, so the reported recall/NDCG and diversity numbers in Tables 3–5 are fitted rather than predicted. The paper needs to specify the split used for reward computation and, if test labels were used, re-run the optimization on the training/validation split before reporting test performance.
[Section 4.3, Fig. 2] The erasure validation is circular. The top-|Δ| features are the entries that the optimization moved to increase diversity; zeroing them undoes the learned intervention, so a decline in diversity and a rise in accuracy is expected by construction. The comparison against least/random erasure does not control for the fact that the erased magnitudes come from the very perturbation that was optimized for diversity. An appropriate control would erase equally large entries from a random perturbation or from a perturbation learned by another method; without such a control, RQ2's conclusion that the selected features are the drivers of diversity is unsupported.

minor comments (6)

[Eq. (13)] The second displayed equality appears to have a typo: the term r_t/n should be r_t, although the final incremental update formula is the standard one.
[Section 4.2 and Table 5] The notation CMBGradientBPRMF-ILAD is inconsistent with the CMB-* naming used elsewhere; please use a consistent form such as CMB^Gradient_ILAD.
[Table 1] The checkmarks and crosses are not visually aligned with the column headers, which makes the comparison matrix difficult to read.
[Section 4.4] The Phones dataset is introduced for the case study but is not listed in Table 2; please specify its size, split, and feature extraction procedure.
[Algorithm 1] The pseudocode uses numpy.linspace; replace it with mathematical notation for reproducibility.
[Section 4.2] The text refers to 'CMBRandom' but the table uses 'CMBBPRMF-Random'; please standardize the name.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'driving feature' explanation and the RQ2 erasure validation are self-referential: they treat the optimized perturbation Δ as independent evidence for the diversity objective that was used to define Δ.

self definitional [Section 3.4 'Meaningful Features Identification as Explanation' (with Eq. 9 in Section 3.2)]
"The values of Δ indicate the influence of item features on the accuracy-diversity trade-off of the recommendation lists generated by the base model g. ... In particular, larger absolute values of Δ correspond to a greater need for the corresponding features to promote greater diversity."

Δ is learned by solving Eq. 9, max_Δ ∥eΨ∥_2^2 − λ1∥Δ∥_1, where eΨ = Diversity(eR_g) is the same diversity measure the paper claims to explain. Consequently, the coordinates of Δ with large magnitude are, by construction, the coordinates the optimizer moved most in order to increase that exact diversity objective. Asserting that these features 'promote greater diversity' restates the optimization objective rather than deriving an independent property from data. The explanation claim is therefore self-definitional: the 'meaningful features' are defined as the entries of Δ that were optimized to maximize the diversity metric, and then that same construction is presented as the discovery that those features drive diversity.
fitted input called prediction [Section 4.3 'Validity Analysis of Explanations (RQ2)', Fig. 2]
"To evaluate whether Δ can discover the meaningful features that improve diversity or balance accuracy and diversity, we follow the widely deployed erasure-based evaluation criterion [16, 66] from Explainable AI. Specifically, we erase the "most meaningful features" from Δ (setting them to 0) and input this modifiedΔ into the pre-trained model g to generate new recommendations. ... Compared with the least/random manners in Fig. 2a, we observe that omitting these meaningful features by the top manner reduces the diversity scores much while increasing the accuracy measures a lot. ..."

The 'most meaningful features' are selected as the largest-magnitude entries of Δ, and Δ itself was fitted to maximize the diversity component of Eq. 9 (or the combined objective in Eq. 14). Erasing the largest entries of a vector that was optimized to increase a function, and then observing that the same function decreases, is an automatic consequence of the optimization; it is not an independent confirmation that those features 'drive' diversity. The least/random erasure conditions merely confirm that near-zero or arbitrary entries matter less, which is also expected from the L1-sparse Δ induced by Eq. 9. The conclusion 'it verifies' overstates what is a consistency check of the optimizer rather than a validation of the explanation.

full rationale

The paper's core optimization framework (Eq. 9) is a direct-maximization method, and comparing its accuracy-diversity trade-off against MMR and DPP is an empirical, non-circular contribution. There is no load-bearing self-citation: the survey [60] is cited only for background taxonomy and not as justification for a central derivation, and no uniqueness theorem is imported from the authors' prior work. However, the explainability chain is self-referential. Δ is learned by maximizing the diversity metric eΨ; Section 3.4 then interprets the magnitude of Δ as evidence that the corresponding features 'promote' diversity, which is a restatement of the objective. Section 4.3's erasure test erases the largest entries of that same learned vector and reports that the same metric drops, which is a sanity check on the optimizer, not an independent validation. The RQ1 diversity gains reported for CMB variants named after the optimized metric are also partly by construction, since the model is directly trained to maximize that metric; the accuracy-diversity trade-off comparison with baselines, however, retains independent content. A separate correctness concern, not circularity, is that Eq. 13 updates each arm's value with the shared reward r_t that depends on the entire joint perturbation Δ, so the per-arm value estimates are not well-defined under the stated independence assumption. Overall, the central explanation claim reduces to the optimization input, giving partial circularity.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the independence and stationarity assumptions of the bandit optimization, the validity of feature-level counterfactual perturbations, and the erasure-based explanation criterion. The hyperparameters A, n_A, epsilon, T, lambda_1, lambda_2 are chosen by hand or tuned, not derived from theory.

free parameters (6)

lambda_1 = 5 for trade-off objective, 0 for single-objective
Controls the L1 sparsity of the perturbation matrix in Eq. 9; affects which features are highlighted as explanations.
lambda_2 = 0.9
Weights accuracy vs diversity in Eq. 14; sensitivity shown in Table 12, but the effect is small.
A = 0.3
Perturbation threshold for arm values in Eq. 11; controls the magnitude of changes and the accuracy-diversity trade-off (Fig 4a).
n_A = 61
Number of arms in the discretized perturbation range; sensitivity in Fig 4b.
epsilon = 0.1
Exploration probability in epsilon-greedy, Eq. 12.
T = 200
Total optimization iterations.

assumptions (3)

ad hoc to paper Different players (entries of Delta) are independent and their per-arm rewards are stationary.
Stated in Section 3.3: 'we assume that different players are independent of each other'; used to justify the per-arm incremental average update in Eq. 13, but the global reward couples all players.
domain assumption Perturbing item latent features and recomputing scores is a valid counterfactual intervention for diversity.
Section 3.2; the paper treats the base model as fixed and assumes feature-level perturbations affect diversity in a way that can be explained by |Delta|.
domain assumption Erasure of the largest |Delta| entries is a valid measure of explanation importance.
Section 4.3, following erasure-based XAI criteria [16, 66]; this assumes feature importance can be read off directly from the learned perturbation magnitude.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Counterfactual Multi-player Bandits for Explainable Recommendation Diversification." pith.science (2026). https://pith.science/paper/YRDPMCIK

@misc{pith2026250521165,
  author       = {Pith},
  title        = {Pith review of: Counterfactual Multi-player Bandits for Explainable Recommendation Diversification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YRDPMCIK}},
  note         = {Machine review of arXiv:2505.21165}
}

read the original abstract

Existing recommender systems tend to prioritize items closely aligned with users' historical interactions, inevitably trapping users in the dilemma of ``filter bubble''. Recent efforts are dedicated to improving the diversity of recommendations. However, they mainly suffer from two major issues: 1) a lack of explainability, making it difficult for the system designers to understand how diverse recommendations are generated, and 2) limitations to specific metrics, with difficulty in enhancing non-differentiable diversity metrics. To this end, we propose a \textbf{C}ounterfactual \textbf{M}ulti-player \textbf{B}andits (CMB) method to deliver explainable recommendation diversification across a wide range of diversity metrics. Leveraging a counterfactual framework, our method identifies the factors influencing diversity outcomes. Meanwhile, we adopt the multi-player bandits to optimize the counterfactual optimization objective, making it adaptable to both differentiable and non-differentiable diversity metrics. Extensive experiments conducted on three real-world datasets demonstrate the applicability, effectiveness, and explainability of the proposed CMB.

Figures

Figures reproduced from arXiv: 2505.21165 by the authors.

**Figure 3.** The feature explanations of CMB-Individual- [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 2.** Comparisons among the accuracy and diversity per [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Comparisons among the accuracy and diversity of [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Shapley Value-driven Data Pruning for Recommender Systems
cs.IR 2025-05 reject novelty 4.0 of 10

SVV prunes recommender training interactions by their estimated Shapley value contribution to autoencoder loss reduction, reporting modest accuracy gains on four datasets but resting on a faulty value-function derivation.

Reference graph

Works this paper leans on

78 extracted references · 67 canonical work pages · cited by 1 Pith paper

[1]

Ilai Bistritz and Nicholas Bambos. 2020. Cooperative multi-player bandit opti- mization. In NeurIPS. 2016–2027

work page 2020
[2]

Ilai Bistritz and Amir Leshem. 2018. Distributed multi-player bandits-a game of thrones approach. In NeurIPS. 7222–7232

work page 2018
[3]

Sébastien Bubeck, Thomas Budzinski, and Mark Sellke. 2021. Cooperative and stochastic multi-player multi-armed bandit: Optimal regret with neither commu- nication nor collisions. In Conference on Learning Theory . 821–822

work page 2021
[4]

Sébastien Bubeck, Nicolo Cesa-Bianchi, et al. 2012. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends ® in Machine Learning 5, 1 (2012), 1–122

2012
[5]

Sébastien Bubeck, Rémi Munos, and Gilles Stoltz. 2009. Pure exploration in multi- armed bandits problems. InInternational conference on Algorithmic learning theory. 23–37

work page 2009
[6]

Jaime Carbonell and Jade Goldstein. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In SIGIR. 335– 336

work page 1998
[7]

Laming Chen, Guoxin Zhang, and Eric Zhou. 2018. Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity. In NeurIPS. 5627–5638

work page 2018
[8]

Wanyu Chen, Pengjie Ren, Fei Cai, Fei Sun, and Maarten de Rijke. 2020. Improving end-to-end sequential recommendations with intent-aware diversification. In CIKM. 175–184

work page 2020

Show all 78 references

[9]

Ziheng Chen, Fabrizio Silvestri, Jia Wang, He Zhu, Hongshik Ahn, and Gabriele Tolomei. 2022. Relax: Reinforcement learning agent explainer for arbitrary predictive models. In CIKM. 252–261

2022
[10]

Heng-Tze Cheng, Levent Koc, Jeremiah Harmsen, Tal Shaked, Tushar Chandra, Hrishi Aradhye, Glen Anderson, Greg Corrado, Wei Chai, Mustafa Ispir, et al
[11]

Peizhe Cheng, Shuaiqiang Wang, Jun Ma, Jiankai Sun, and Hui Xiong. 2017. Learning to recommend accurate and diverse items. In WWW. 183–192

2017
[12]

Charles LA Clarke, Maheedhar Kolla, Gordon V Cormack, Olga Vechtomova, Azin Ashkan, Stefan Büttcher, and Ian MacKinnon. 2008. Novelty and diversity in information retrieval evaluation. In SIGIR. 659–666

2008
[13]

Qinxu Ding, Yong Liu, Chunyan Miao, Fei Cheng, and Haihong Tang. 2021. A hybrid bandit framework for diversified recommendation. In AAAI. 4036–4044

2021
[14]

Lu Gan, Diana Nurbakova, Léa Laporte, and Sylvie Calabretto. 2020. Enhancing recommendation diversity using determinantal point processes on knowledge graphs. In SIGIR. 2001–2004

2020
[15]

Mouzhi Ge, Carla Delgado-Battenfeld, and Dietmar Jannach. 2010. Beyond accuracy: evaluating recommender systems by coverage and serendipity. In RecSys. 257–260

2010
[16]

Yingqiang Ge, Juntao Tan, Yan Zhu, Yinglong Xia, Jiebo Luo, Shuchang Liu, Zuohui Fu, Shijie Geng, Zelong Li, and Yongfeng Zhang. 2022. Explainable fairness in recommendation. In SIGIR. 681–691

2022
[17]

Azin Ghazimatin, Oana Balalau, Rishiraj Saha Roy, and Gerhard Weikum. 2020. PRINCE: Provider-side interpretability with counterfactual explanations in rec- ommender systems. In WSDM. 196–204

2020
[18]

Wenwen Gong, Xuyun Zhang, Yifei Chen, Qiang He, Amin Beheshti, Xiaolong Xu, Chao Yan, and Lianyong Qi. 2022. DAWAR: Diversity-aware Web APIs Recommendation for Mashup Creation based on Correlation Graph. In SIGIR. 395–404

2022
[19]

Jungkyu Han and Hayato Yamana. 2019. Geographic diversification of recom- mended pois in frequently visited areas. TOIS 38, 1 (2019), 1–39

2019
[20]

F Maxwell Harper and Joseph A Konstan. 2015. The movielens datasets: History and context. TIIS 5, 4 (2015), 1–19

2015
[21]

Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yong-Dong Zhang, and Meng Wang. 2020. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation. In SIGIR. 639–648

2020
[22]

Yifan He, Haitao Zou, Hualong Yu, Qi Wang, and Shang Gao. 2019. Diversity- Aware Recommendation by User Interest Domain Coverage Maximization. In ICDM. 1084–1089

2019
[23]

Jonathan L Herlocker, Joseph A Konstan, Loren G Terveen, and John T Riedl
[24]

Yanhua Huang, Weikun Wang, Lei Zhang, and Ruiwen Xu. 2021. Sliding Spectrum Decomposition for Diversified Recommendation. In KDD. 3041–3049

2021
[25]

Zohar Karnin, Tomer Koren, and Oren Somekh. 2013. Almost optimal exploration in multi-armed bandits. In ICML. 1238–1246

2013
[26]

Emilie Kaufmann, Olivier Cappé, and Aurélien Garivier. 2012. On Bayesian upper confidence bounds for bandit problems. In Artificial intelligence and statistics . 592–600

2012
[27]

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

2014 arXiv
[28]

Bell, and Chris Volinsky

Yehuda Koren, Robert M. Bell, and Chris Volinsky. 2009. Matrix Factorization Techniques for Recommender Systems. Computer 42, 8 (2009), 30–37

2009
[29]

Matevž Kunaver and Tomaž Požrl. 2017. Diversity in recommender systems–A survey. Knowledge-based systems 123 (2017), 154–162

2017
[30]

Hyokmin Kwon, Jaeho Han, and Kyungsik Han. 2020. ART (Attractive Rec- ommendation Tailor) How the Diversity of Product Recommendations Affects Customer Purchase Preference in Fashion Industry?. In CIKM. 2573–2580

2020
[31]

Chang Li, Haoyun Feng, and Maarten de Rijke. 2020. Cascading hybrid bandits: Online learning to rank for relevance and diversity. In RecSys. 33–42

2020
[32]

Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. In WWW. 661– 670

2010
[33]

Shuang Li, Yuezhi Zhou, Di Zhang, Yaoxue Zhang, and Xiang Lan. 2017. Learning to diversify recommendations based on matrix factorization. In DASC/PiCom/DataCom/CyberSciTech. 68–74

2017
[34]

Xueqi Li, Wenjun Jiang, Weiguang Chen, Jie Wu, Guojun Wang, and Kenli Li. 2020. Directional and explainable serendipity recommendation. In WWW. 122–132

2020
[35]

Yichao Lu, Ruihai Dong, and Barry Smyth. 2018. Coevolutionary recommendation model: Mutual learning between ratings and reviews. In WWW. 773–782

2018
[36]

Stella Maropaki, Sean Chester, Christos Doulkeridis, and Kjetil Nørvåg. 2020. Diversifying Top-k Point-of-Interest Queries via Collective Social Reach. InCIKM. 2149–2152

2020
[37]

Robert K Merton. 1968. The Matthew effect in science: The reward and commu- nication systems of science are considered. Science 159, 3810 (1968), 56–63

1968
[38]

Jianmo Ni, Jiacheng Li, and Julian J. McAuley. 2019. Justifying Recommendations using Distantly-Labeled Reviews and Fine-Grained Aspects. In EMNLP-IJCNLP. 188–197

2019
[39]

Javier Parapar and Filip Radlinski. 2021. Diverse user preference elicitation with multi-armed bandits. In WSDM. 130–138

2021
[40]

Eli Pariser. 2011. The filter bubble: How the new personalized web is changing what we read and how we think

2011
[41]

Bibek Paudel, Thilo Haas, and Abraham Bernstein. 2017. Fewer flops at the top: Accuracy, diversity, and regularization in two-class collaborative filtering. In RecSys. 215–223

2017
[42]

Lijing Qin, Shouyuan Chen, and Xiaoyan Zhu. 2014. Contextual combinatorial bandit and its application on diversified online recommendation. In SIAM. 461– 469

2014
[43]

Xubo Qin, Zhicheng Dou, and Ji-Rong Wen. 2020. Diversifying search results using self-attention network. In CIKM. 1265–1274

2020
[44]

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
[45]

Why should i trust you?

Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why should i trust you?" Explaining the predictions of any classifier. In KDD. 1135–1144

2016
[46]

Rodrygo LT Santos, Craig Macdonald, and Iadh Ounis. 2010. Exploiting query reformulations for web search result diversification. In WWW. 881–890

2010
[47]

Javier Sanz-Cruzado and Pablo Castells. 2018. Enhancing structural diversity in social networks by recommending weak ties. In RecSys. 233–241

2018
[48]

Xiaoyu Shi, Quanliang Liu, Hong Xie, Di Wu, Bo Peng, MingSheng Shang, and Defu Lian. 2023. Relieving popularity bias in interactive recommendation: A diversity-novelty-aware reinforcement learning approach. TOIS 42, 2 (2023), 1–30

2023
[49]

Jaspreet Singh and Avishek Anand. 2019. Exs: Explainable search using local model agnostic interpretability. In WSDM. 770–773

2019
[50]

Aleksandrs Slivkins et al. 2019. Introduction to multi-armed bandits. Foundations and Trends® in Machine Learning 12, 1-2 (2019), 1–286

2019
[51]

Dusan Stamenkovic, Alexandros Karatzoglou, Ioannis Arapakis, Xin Xin, and Kleomenis Katevas. 2022. Choosing the Best of Both Worlds: Diverse and Novel Recommendations through Multi-Objective Reinforcement Learning. In WSDM. 957–965

2022
[52]

Juntao Tan, Shuyuan Xu, Yingqiang Ge, Yunqi Li, Xu Chen, and Yongfeng Zhang
[53]

Kosetsu Tsukuda and Masataka Goto. 2019. DualDiv: diversifying items and explanation styles in explainable hybrid recommendation. In RecSys. 398–402

2019
[54]

Izzatul Umami and Lailia Rahmawati. 2021. Comparing Epsilon greedy and Thompson sampling model for multi-armed bandit algorithm on marketing dataset. Journal of Applied Data Sciences 2, 2 (2021)

2021
[55]

Saul Vargas, Pablo Castells, and David Vallet. 2011. Intent-oriented diversity in recommender systems. In SIGIR. 1211–1212

2011
[56]

Sandra Wachter, Brent Mittelstadt, and Chris Russell. 2017. Counterfactual explanations without opening the black box: Automated decisions and the GDPR. Harv. JL & Tech. 31 (2017), 841

2017
[57]

Xiting Wang, Yiru Chen, Jie Yang, Le Wu, Zhengtao Wu, and Xing Xie. 2018. A reinforcement learning framework for explainable recommendation. In ICDM. 587–596

2018
[58]

Jacek Wasilewski and Neil Hurley. 2016. Incorporating diversity in a learning to rank recommender system. In FLAIRS. 572–578

2016
[59]

Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H Chi, and Jennifer Gillenwater. 2018. Practical diversified recommendations on youtube with determinantal point processes. In CIKM. 2165–2173

2018
[60]

Haolun Wu, Yansen Zhang, Chen Ma, Fuyuan Lyu, Bowei He, Bhaskar Mitra, and Xue Liu. 2024. Result Diversification in Search and Recommendation: A Survey. TKDE (2024)

2024
[61]

Libing Wu, Cong Quan, Chenliang Li, Qian Wang, Bolong Zheng, and Xiangyang Luo. 2019. A context-aware user-item representation learning for item recom- mendation. TOIS 37, 2 (2019), 1–29

2019
[62]

Kun Xiong, Wenwen Ye, Xu Chen, Yongfeng Zhang, Wayne Xin Zhao, Binbin Hu, Zhiqiang Zhang, and Jun Zhou. 2021. Counterfactual review-based recom- mendation. In CIKM. 2231–2240

2021
[63]

Liangwei Yang, Shengjie Wang, Yunzhe Tao, Jiankai Sun, Xiaolong Liu, Philip S Yu, and Taiqing Wang. 2023. DGRec: Graph Neural Network for Recommendation with Diversified Embedding Generation. In WSDM. 661–669

2023
[64]

Rui Ye, Yuqing Hou, Te Lei, Yunxing Zhang, Qing Zhang, Jiale Guo, Huaiwen Wu, and Hengliang Luo. 2021. Dynamic graph construction for improving diversity of recommendation. In RecSys. 651–655

2021
[65]

Hai-Tao Yu. 2022. Optimize What You Evaluate With: Search Result Diversifica- tion Based on Metric Optimization. In AAAI. 10399–10407

2022
[66]

Jaakkola

Mo Yu, Shiyu Chang, Yang Zhang, and Tommi S. Jaakkola. 2019. Rethinking Cooperative Rationalization: Introspective Extraction and Complement Control. In EMNLP-IJCNLP. 4092–4101

2019
[67]

Mi Zhang and Neil Hurley. 2008. Avoiding monotony: improving the diversity of recommendation lists. In RecSys. 123–130

2008
[68]

Yongfeng Zhang, Xu Chen, et al. 2020. Explainable recommendation: A survey and new perspectives. Foundations and Trends ® in Information Retrieval 14, 1 (2020), 1–101

2020
[69]

Yansen Zhang, Chenhao Hu, Genan Dai, Weiyang Kong, and Yubao Liu. 2021. Self- adaptive Graph Neural Networks for Personalized Sequential Recommendation. In ICONIP. 608–619

2021
[70]

Yongfeng Zhang, Guokun Lai, Min Zhang, Yi Zhang, Yiqun Liu, and Shaoping Ma. 2014. Explicit factor models for explainable recommendation based on phrase-level sentiment analysis. In SIGIR. 83–92

2014
[71]

Yansen Zhang, Xiaokun Zhang, Ziqiang Cui, and Chen Ma. 2025. Shapley Value- driven Data Pruning for Recommender Systems. In KDD

2025
[72]

Guanjie Zheng, Fuzheng Zhang, Zihan Zheng, Yang Xiang, Nicholas Jing Yuan, Xing Xie, and Zhenhui Li. 2018. DRN: A deep reinforcement learning framework for news recommendation. In WWW. 167–176

2018
[73]

Yu Zheng, Chen Gao, Liang Chen, Depeng Jin, and Yong Li. 2021. DGCN: Diversi- fied Recommendation with Graph Convolutional Networks. In WWW. 401–412

2021
[74]

Jianghong Zhou, Eugene Agichtein, and Surya Kallumadi. 2020. Diversifying Multi-aspect Search Results Using Simpson’s Diversity Index. In CIKM. 2345– 2348. A Experiments A.1 Experiment Setup A.1.1 Datasets. • MovieLens [20]: This dataset contains user ratings on MovieLens web ...

2020
[2004]

TOIS 22, 1 (2004), 5–53

Evaluating collaborative filtering recommender systems. TOIS 22, 1 (2004), 5–53

2004
[2009]

BPR: Bayesian Personalized Ranking from Implicit Feedback. In UAI. 452– 461. Counterfactual Multi-player Bandits for Explainable Recommendation Diversification Conference acronym ’XX, June 03–05, 2018, Woodstock, NY

2018
[2016]

In Proceedings of the 1st workshop on deep learning for recommender systems

Wide & deep learning for recommender systems. In Proceedings of the 1st workshop on deep learning for recommender systems . 7–10
[2021]

Counterfactual explainable recommendation. In CIKM. 1784–1793

Pith tools

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

[1] [1]

Ilai Bistritz and Nicholas Bambos. 2020. Cooperative multi-player bandit opti- mization. In NeurIPS. 2016–2027

work page 2020

[2] [2]

Ilai Bistritz and Amir Leshem. 2018. Distributed multi-player bandits-a game of thrones approach. In NeurIPS. 7222–7232

work page 2018

[3] [3]

Sébastien Bubeck, Thomas Budzinski, and Mark Sellke. 2021. Cooperative and stochastic multi-player multi-armed bandit: Optimal regret with neither commu- nication nor collisions. In Conference on Learning Theory . 821–822

work page 2021

[4] [4]

Sébastien Bubeck, Nicolo Cesa-Bianchi, et al. 2012. Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Foundations and Trends ® in Machine Learning 5, 1 (2012), 1–122

2012

[5] [5]

Sébastien Bubeck, Rémi Munos, and Gilles Stoltz. 2009. Pure exploration in multi- armed bandits problems. InInternational conference on Algorithmic learning theory. 23–37

work page 2009

[6] [6]

Jaime Carbonell and Jade Goldstein. 1998. The use of MMR, diversity-based reranking for reordering documents and producing summaries. In SIGIR. 335– 336

work page 1998

[7] [7]

Laming Chen, Guoxin Zhang, and Eric Zhou. 2018. Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity. In NeurIPS. 5627–5638

work page 2018

[8] [8]

Wanyu Chen, Pengjie Ren, Fei Cai, Fei Sun, and Maarten de Rijke. 2020. Improving end-to-end sequential recommendations with intent-aware diversification. In CIKM. 175–184

work page 2020