Pith. sign in

REVIEW 4 major objections 3 minor 32 references

Leveraging the Power of Conversations: Optimal Key Term Selection in Conversational Contextual Bandits

T0 review · 4 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Conversational contextual bandits can reach near-optimal regret $O(\sqrt{dT\log T})$ by asking noisy, uncertainty-targeted questions, with a matching $\Omega(\sqrt{dT})$ lower bound for any policy that asks at most one question per round.

desk verdict The headline regret bounds for CLiSK and CLiSK-ME rest on a proof that changes the reward model midstream, and the CLiME analysis has a load-bearing gap; the ideas are novel but the paper is not ready as written. read the letter →

arxiv 2505.21393 v1 pith:BAUA2O5E submitted 2025-05-27 cs.LG cs.AI

classification cs.LGcs.AI
keywords conversationalrecommendersystemscontextualbanditspreferencelearningregretminimizationsmoothedanalysiskeytermselectionminimaxoptimalityonline
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

Conversational recommender systems can learn a user's tastes faster by asking about genres or keywords, but earlier algorithms for this problem were no better, in the worst case, than systems that never ask. This paper claims to close that gap. It introduces three algorithms, CLiSK, CLiME, and CLiSK-ME, built on two ideas: add a small amount of noise to key-term vectors before greedily choosing which term to ask about, and start conversations only when the uncertainty ellipse of the learned preference is large in some direction. All three achieve cumulative regret of order $O(\sqrt{dT\log T})$ (with an extra dimension-only additive term for CLiSK), the paper proves a matching lower bound $\Omega(\sqrt{dT})$, and experiments report at least a 14.6% improvement over existing methods. If correct, the paper settles the open question of whether conversational feedback can improve the worst-case regret rate of linear contextual bandits.

What carries the argument

The two load-bearing mechanisms are smoothed key term contexts and a minimum-eigenvalue conversation trigger. A smoothed key term context is a key-term vector $\tilde{\mathbf{x}}_k$ plus an independent truncated Gaussian vector $\boldsymbol{\varepsilon}_k\sim N(0,\rho^2 I_d)$ bounded by $R$; it does the work of forcing the selected query's outer product to have uniform positive variance in every direction, so that even greedy key-term selection explores the whole feature space. The minimum-eigenvalue trigger, used by CLiME and CLiSK-ME, diagonalizes the covariance matrix $M_t$, finds eigenvectors with $\lambda_{v_i}<\alpha_t$, and schedules enough aligned key-term queries to push $\lambda_{\min}(M_t)$ back above $\alpha_t$. It is what makes conversation timing adaptive rather than fixed by a schedule. The proof machinery around these objects, including matrix Chernoff bounds, determinant-trace bounds, and a KL-divergence lower-bound construction, turns the eigenvalue growth directly into regret bounds.

What would settle it

Run CLiME on a synthetic problem whose key-term vectors all lie in a strict subspace of $\mathbb{R}^d$, with the optimal arm outside that subspace, and record $\lambda_{\min}(M_t)$: if it no longer stays above the threshold $\alpha_t$, the eigenvalue-boosting argument fails exactly where Assumption 3 is violated. For CLiSK, give users binary feedback on blurred genres as well as on the original genres; if observed feedback tracks $\tilde{\mathbf{x}}_k^\top \boldsymbol{\theta}^*$ rather than $(\tilde{\mathbf{x}}_k+\boldsymbol{\varepsilon}_k)^\top \boldsymbol{\theta}^*$, the estimator is biased and cumulative regret should visibly grow.

Watch

Extended reading notes

Core claim

The central discovery is that conversational feedback stops being redundant once key-term queries are forced to be informative in every direction. CLiSK replaces each key-term vector $\tilde{\mathbf{x}}_k$ with a truncated Gaussian perturbation $\tilde{\tilde{\mathbf{x}}}_k = \tilde{\mathbf{x}}_k + \boldsymbol{\varepsilon}_k$, so the greedy pick $\arg\max_k \tilde{\tilde{\mathbf{x}}}_k^\top \boldsymbol{\theta}_t$ has expected outer product with minimum eigenvalue at least $c_1\rho^2/\log|\mathcal{K}|$. This makes the covariance matrix of key-term feedback grow at rate $\lambda_{\mathcal{K}} b t$, which bounds the UCB radius by $\sqrt{2/(\lambda_{\mathcal{K}} b t)}$ and yields $O(\sqrt{dT\log T}+d)$ regret. CLiME uses an adaptive trigger: it diagonalizes the covariance matrix and, whenever an eigenvalue drops below $\alpha_t$, schedules $\lceil(\alpha_t - \lambda_{v_i})/c_0^2\rceil$ conversations with a key term aligned with that eigenvector, proving $\lambda_{\min}(M_t)\ge \alpha_t$ and bounded total conversations $O(\alpha d t/c_0^2)$. The combination CLiSK-ME inherits both properties. The paper also constructs a lower bound $\Omega(\sqrt{dT})$ by splitting the $d$-dimensional problem into independent lower-dimensional instances, showing that the dependence on $T$ cannot be improved; this is presented as the first $\tilde O(\sqrt{dT})$ result in the conversational bandit literature.

Load-bearing premise

The load-bearing premise is that the set of key terms is rich enough that for any direction a user's taste could point, some key term points almost exactly that way, and that a user's answer to a slightly blended key term is still exactly a straight-line function of the blended term; if either fails, the claimed regret bounds no longer follow.

Editorial extensions

If this is right

  • The regret rate of conversational linear bandits improves from $O(d\sqrt{T\log T})$ to $O(\sqrt{dT\log T})$, so asking well-chosen key-term questions now provably helps in the worst case.
  • Conversation timing can be made adaptive without hurting the guarantee: continuous, fixed-interval, and exponential-phase uncertainty checks all keep the same regret bound, letting a system stop interrupting users once preferences are well estimated.
  • The matching lower bound $\Omega(\sqrt{dT})$ applies to every policy that asks at most one key-term question per round, so no algorithm of that restricted form will asymptotically beat CLiME or CLiSK-ME.
  • On synthetic, MovieLens, Yelp, and Last.fm data, the three algorithms reduce cumulative regret by more than 14.6% relative to the best prior baselines, matching the theoretical ordering.

Reading between the lines

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

  • An unstated assumption is that a user's answer to a blended key term is linear in the blended vector. If human responses are actually linear in the original key-term vector, the CLiSK and CLiSK-ME updates would be biased; a controlled study comparing responses on original versus blended genre prompts would settle this.
  • Assumption 3 asks a finite vocabulary to nearly cover the unit sphere, which is hard to satisfy in high dimension with sparse category systems. If the coverage constant $c_0$ shrinks with $d$, the conversation-count bound $O(\alpha d t/c_0^2)$ and the eigenvalue boost degrade, so the practical gain likely depends on how structured real key-term sets are.
  • The lower bound assumes at most one key-term query per round. A natural open direction is whether multiple simultaneous key-term queries per round, or a batched conversation, could beat $\Omega(\sqrt{dT})$ or whether the bound can be extended to that setting.
  • The uncertainty-checking functions imply a deployable design rule: ask the user only when some eigen-direction of the confidence ellipsoid is too thin. This can be evaluated directly with engagement or interruption-cost metrics, not only with regret.
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.

Referee Report

4 major / 3 minor

Summary. The paper studies conversational contextual bandits, where in addition to arm-level rewards the agent may query the user about key terms. It proposes three algorithms: CLiSK (smoothed key-term contexts with fixed-frequency conversations), CLiME (adaptive conversation initiation based on eigenvalue uncertainty), and CLiSK-ME (a combination). For all three it claims regret upper bounds of O(√(dT log T)) (with an extra additive d for CLiSK), and it claims a matching lower bound Ω(√(dT)) for conversational bandits. The paper also reports experiments on synthetic and real-world datasets showing more than 14.6% regret improvement over baselines.

Significance. If the theoretical results were correct, the paper would be significant: it would be the first to obtain Õ(√(dT)) regret in conversational contextual bandits, improving on the earlier O(d√T log T) bounds, and it would establish near-minimax optimality. The algorithmic ideas—adding noise to key-term contexts and adaptively triggering conversations based on uncertainty—are interesting and the experimental section is extensive, with a public code link. However, the central theoretical claims are not supported: there is an internal inconsistency between the stated feedback model and the proof of the CLiSK regret bound, a load-bearing invalid step in the CLiME eigenvalue argument, and an apparent contradiction between the paper's lower-bound lemma and its own upper bounds. Because these issues affect the main claims of the paper, the current manuscript cannot be accepted.

major comments (4)
  1. [Section 2, Algorithm 1, and Appendix A.3] The key-term feedback model is inconsistent with the ridge-regression update and with the proof of Lemma 1. Section 2 defines the key-term reward as linear in the original key-term context: \tilde r_{k,t} = \tilde x_k^\top \theta^* + \tilde \eta_t. Algorithm 1, however, smooths contexts to \tilde{\tilde x}_k = \tilde x_k + \varepsilon_k, selects k based on \tilde{\tilde x}_k, and updates b_t using \tilde r_{k,t} \tilde{\tilde x}_k. In Appendix A.3 the proof of Lemma 1 silently replaces the reward model, writing \tilde r_{k,s} = \tilde{\tilde x}_k^\top \theta^* + \tilde \eta_s. Under the model stated in Section 2, the conditional expectation of the key-term contribution to b_t is \tilde{\tilde x}_k \tilde x_k^\top \theta^*, not \tilde{\tilde x}_k \tilde{\tilde x}_k^\top \theta^*. The discrepancy is \tilde{\tilde x}_k \varepsilon_k^\top \theta^*, which is not zero-mean conditional on the selection event because k = argmax_j \tilde{\tilde x}_j^\top \theta_t depends on \varepsilon_k. The ridge estimator is therefore biased, the self-normalized martingale bound used in Equation (3) does not apply, and Theorem 1 and Corollary 1 are unsupported.
  2. [Appendix A.7, proof of Lemma 5] The proof of the key eigenvalue-boosting claim contains an invalid Loewner inequality. The proof writes \tilde x_k = \sum_i \gamma_i v_i, sets z_j = \sum_{i \neq j} \gamma_i v_i, and then claims \tilde x_k \tilde x_k^\top = (\gamma_j v_j + z_j)(\gamma_j v_j + z_j)^\top = \gamma_j^2 v_j v_j^\top + z_j z_j^\top. This expansion drops the cross terms \gamma_j(v_j z_j^\top + z_j v_j^\top), which are not zero. Additionally, the assertion that \gamma_j \ge c_0 for every coordinate j does not follow from Algorithm 2's selection rule, which only guarantees |\tilde x_k^\top v_i| \ge c_0 for the particular eigenvector v_i being targeted. A rank-one matrix \tilde x_k \tilde x_k^\top does not dominate c_0^2 v_i v_i^\top in the Loewner order merely from |\tilde x_k^\top v_i| \ge c_0. Consequently Equation (13), and hence \lambda_{\min}(M_t) \ge \alpha_t, is not established. Lemma 6, Theorem 2, and Corollary 1 all rely on this step, so the CLiME and CLiSK-ME upper bounds are not supported as written.
  3. [Appendix A.11, Theorem 3 and Lemma 8] The lower-bound section appears to contradict the paper's own upper bounds. Lemma 8 concludes an expected regret lower bound of \Omega(d\sqrt{T}) for the constructed instance. For d sufficiently large relative to \log T, this is strictly larger than the claimed upper bounds O(\sqrt{dT\log T}) of Theorems 1 and 2. At least one of the lower-bound or upper-bound arguments must therefore be false, and the claim of near-minimax optimality is not coherent. The proof also uses A = K = [-1,1]^d, which is not a finite arm/key-term set as assumed in Section 2, and the displayed equality expressing regret as a sum of sign-mismatch indicators is not correct as written because suboptimal choices with the same sign but small magnitude also contribute to regret. These points need to be reconciled before the lower-bound claim can be accepted.
  4. [Assumption 3] Assumption 3 is load-bearing for both CLiME and CLiSK-ME and is presented as mild, but it is very strong in high dimensions. Requiring that for every unit direction x there exists a key term with |\tilde x_k^\top x| \ge c_0 for a constant c_0 close to 1 implies that a finite key-term set nearly covers the unit sphere, which typically requires |K| exponential in d. Since the scheduling counts n_k and the trace/determinant arguments in Lemma 5 depend critically on c_0, the theorem's applicability to the regime d \gg \log|K|, which is the regime promised by the stated regret bound, is not established.
minor comments (3)
  1. [Appendix A.11] In the proof of Lemma 8, the displayed calculation after the KL divergence contains a typo: it reads "⟨A_t, θ−θ′⟩ = ⟨A_t, θ−θ′⟩ = 4/T" rather than the intended squared-inner-product statement.
  2. [Definition 1 and Lemma 3] Lemma 2 states a lower bound on E[\tilde{\tilde x}_k \tilde{\tilde x}_k^\top] for the selected k, but the proof conditions on the current \theta_t; the notation should specify that the expectation is conditional on the history and clarify how the matrix Chernoff bound in Lemma 3 handles the dependence of the selected key terms across rounds.
  3. [Section 5.2] The figures report cumulative regret normalized to 1e3 without stating whether the plotted quantity is the average over the 20 trials; the caption should clarify the error bars and the normalization.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CLiSK, CLiME, and CLiSK-ME regret bounds rest on external concentration inequalities and a standard information-theoretic lower bound, not on fitted parameters or self-citation chains.

full rationale

The derivation chain is self-contained in the direction that matters for circularity. CLiSK's regret bound is built from the self-normalized martingale inequality of Abbasi-Yadkori et al. [1], a determinant-trace bound, and eigenvalue lower bounds for smoothed selected key terms whose proofs cite Tropp [24] and Sivakumar et al. [22] for the anti-concentration machinery. CLiME uses the same external concentration framework plus its explicit eigenvalue-boosting schedule; the number of conversations is bounded from the algorithm's own stopping rule, not from the target regret bound. The lower bound in Theorem 3 / Appendix A.11 is a standard Bretagnolle-Huber / KL-divergence argument with no fitted parameters. Self-citations to Zhang et al. [29] and Wang et al. [25] set the problem formulation and the linear query-frequency schedule, but the load-bearing inequalities are external and independently checkable, and no uniqueness theorem from the same authors is invoked to forbid alternatives. I therefore find no step in which a claimed prediction is equivalent to its input by construction, and no fitted parameter is renamed as a prediction. A separate concern exists: the proof of Lemma 1 writes the key-term reward as linear in the smoothed context while Section 2 defines it as linear in the original context; that is a model-mismatch or correctness issue, not a circularity, because the bound is not derived from the stated model by an identity but under an unstated alternative model. The stray PhaseElimination code block in the extracted text makes no assertion about this paper's algorithms and does not affect the derivation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central regret bounds rest on two explicit domain assumptions (normalization, sub-Gaussian noise), one very strong diversity assumption on the key term set, and an implicit reward-model assumption for smoothed contexts that is inconsistent with the formal model. The algorithms have several hand-chosen hyperparameters (perturbation level, truncation, exploration control, regularization) that are not fitted to data but are also not fully specified for experiments.

free parameters (4)
  • rho^2 = 1 (experiments)
    Gaussian perturbation variance in CLiSK and CLiSK-ME; hand-chosen, ablation shows regret is insensitive to it between 0.1 and 3.
  • R = 1 (experiments)
    Truncation bound for the perturbation; ablation shows regret decreases as R grows and stabilizes only when R > 2, so the chosen value is not the best hand-tuned choice.
  • alpha = not reported
    Exploration control in CLiME; appears in the regret guarantee and conversation count but no value or tuning procedure is given in the evaluation.
  • lambda = not reported
    Ridge regularization parameter; standard but no sensitivity study.
assumptions (4)
  • domain assumption Feature vectors of arms and key terms are normalized and ||theta*||<=1 (Assumption 1)
    Standard in linear bandits; restricts problem but is stated.
  • domain assumption Noise terms are conditionally 1-sub-Gaussian (Assumption 2)
    Standard noise model enabling self-normalized bounds.
  • domain assumption Key term set is rich: for every unit x there exists k with |tilde x_k^T x| >= c0, c0 close to 1 (Assumption 3)
    Load-bearing for CLiME; without it, uncertain directions cannot be queried and the eigenvalue maintenance argument fails. It is very strong for finite K in high dimension.
  • ad hoc to paper User feedback on a smoothed key term is generated by the smoothed vector tilde tilde x_k (implicit in Lemma 1 proof)
    Nowhere stated in Section 2; contradicts the formal reward model and invalidates the CLiSK/CLiSK-ME concentration arguments as written.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging the Power of Conversations: Optimal Key Term Selection in Conversational Contextual Bandits." pith.science (2026). https://pith.science/paper/BAUA2O5E

@misc{pith2026250521393,
  author       = {Pith},
  title        = {Pith review of: Leveraging the Power of Conversations: Optimal Key Term Selection in Conversational Contextual Bandits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BAUA2O5E}},
  note         = {Machine review of arXiv:2505.21393}
}
abstract

Conversational recommender systems proactively query users with relevant "key terms" and leverage the feedback to elicit users' preferences for personalized recommendations. Conversational contextual bandits, a prevalent approach in this domain, aim to optimize preference learning by balancing exploitation and exploration. However, several limitations hinder their effectiveness in real-world scenarios. First, existing algorithms employ key term selection strategies with insufficient exploration, often failing to thoroughly probe users' preferences and resulting in suboptimal preference estimation. Second, current algorithms typically rely on deterministic rules to initiate conversations, causing unnecessary interactions when preferences are well-understood and missed opportunities when preferences are uncertain. To address these limitations, we propose three novel algorithms: CLiSK, CLiME, and CLiSK-ME. CLiSK introduces smoothed key term contexts to enhance exploration in preference learning, CLiME adaptively initiates conversations based on preference uncertainty, and CLiSK-ME integrates both techniques. We theoretically prove that all three algorithms achieve a tighter regret upper bound of $O(\sqrt{dT\log{T}})$ with respect to the time horizon $T$, improving upon existing methods. Additionally, we provide a matching lower bound $\Omega(\sqrt{dT})$ for conversational bandits, demonstrating that our algorithms are nearly minimax optimal. Extensive evaluations on both synthetic and real-world datasets show that our approaches achieve at least a 14.6% improvement in cumulative regret.

Figures

Figures reproduced from arXiv: 2505.21393 by the authors.

Figure 1
Figure 1. Illustration of conversational recommendation by [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of cumulative regret where CLiME and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Comparison of cumulative regret where CLiME and [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Comparison of estimation precision where CLiME [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Number of conversations initiated by deterministic [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Effect of the truncation limit 𝑅. 6 Related Work Our research is closely aligned with studies on conversational con￾textual bandits, particularly focusing on the problem of key term selection within this framework. Contextual bandits serve as a fundamental framework fo…
Figure 8
Figure 8. Figure 8: Comparison of estimation precision where CLiME [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: Comparison of estimation precision where CLiME [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 28 canonical work pages

  1. [1]

    Yasin Abbasi-Yadkori, Dávid Pál, and Csaba Szepesvári. 2011. Improved algo- rithms for linear stochastic bandits. Advances in neural information processing systems 24 (2011)

  2. [2]

    Shipra Agrawal and Navin Goyal. 2012. Analysis of thompson sampling for the multi-armed bandit problem. In Conference on learning theory . JMLR Workshop and Conference Proceedings, 39–1

  3. [3]

    Bretagnolle and C

    J. Bretagnolle and C. Huber. 1978. Estimation des densités : Risque minimax. In Séminaire de Probabilités XII, C. Dellacherie, P. A. Meyer, and M. Weil (Eds.). Springer Berlin Heidelberg, Berlin, Heidelberg, 342–363

  4. [4]

    Ivan Cantador, Peter Brusilovsky, and Tsvi Kuflik. 2011. Second Workshop on Information Heterogeneity and Fusion in Recommender Systems (HetRec2011). In Proceedings of the Fifth ACM Conference on Recommender Systems (RecSys ’11) . 387–388

  5. [5]

    Konstantina Christakopoulou, Filip Radlinski, and Katja Hofmann. 2016. Towards conversational recommender systems. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining . 815–824

  6. [6]

    Wei Chu, Lihong Li, Lev Reyzin, and Robert Schapire. 2011. Contextual bandits with linear payoff functions. In Proceedings of the Fourteenth International Con- ference on Artificial Intelligence and Statistics . JMLR Workshop and Conference Proceedings, 208–214

  7. [7]

    Xiangxiang Dai, Zhiyong Wang, Jize Xie, Xutong Liu, and John CS Lui. 2024. Conversational Recommendation with Online Learning and Clustering on Mis- specified Users. IEEE Transactions on Knowledge and Data Engineering 36, 12 (2024), 7825–7838

  8. [8]

    Xiangxiang Dai, Zhiyong Wang, Jize Xie, Tong Yu, and John CS Lui. 2024. Online Learning and Detecting Corrupted Users for Conversational Recommendation Systems. IEEE Transactions on Knowledge and Data Engineering 36, 12 (2024), 8939–8953

Show all 32 references
  1. [9]

    Xiangxiang Dai, Yuejin Xie, Maoli Liu, Xuchuang Wang, Zhuohua Li, Huanyu Wang, and John C. S. Lui. 2025. Multi-Agent Conversational Online Learning for Adaptive LLM Response Identification. arXiv:2501.01849 [cs.HC]

  2. [10]

    Yi Gai, Bhaskar Krishnamachari, and Rahul Jain. 2012. Combinatorial network optimization with unknown variables: Multi-armed bandits with linear rewards and individual observations. IEEE/ACM Transactions on Networking 20, 5 (2012), 1466–1478

  3. [11]

    Chongming Gao, Wenqiang Lei, Xiangnan He, Maarten de Rijke, and Tat-Seng Chua. 2021. Advances and challenges in conversational recommender systems: A survey. AI open 2 (2021), 100–126

  4. [12]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. 2015. The MovieLens Datasets: History and Context. ACM Trans. Interact. Intell. Syst. 5, 4, Article 19 (dec 2015), 19 pages

  5. [13]

    Sampath Kannan, Jamie H Morgenstern, Aaron Roth, Bo Waggoner, and Zhi- wei Steven Wu. 2018. A smoothed analysis of the greedy algorithm for the linear contextual bandit problem. Advances in neural information processing systems 31 (2018)

  6. [14]

    Tor Lattimore and Csaba Szepesvári. 2020. Bandit Algorithms. Cambridge Uni- versity Press

  7. [15]

    Lihong Li, Wei Chu, John Langford, and Robert E Schapire. 2010. A contextual- bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web . 661–670

  8. [16]

    Shijun Li, Wenqiang Lei, Qingyun Wu, Xiangnan He, Peng Jiang, and Tat-Seng Chua. 2021. Seamlessly unifying attributes and items: Conversational recommen- dation for cold-start users. ACM Transactions on Information Systems (TOIS) 39, 4 (2021), 1–29

  9. [17]

    Zhuohua Li, Maoli Liu, Xiangxiang Dai, and John C.S. Lui. 2025. Demystify- ing Online Clustering of Bandits: Enhanced Exploration Under Stochastic and Smoothed Adversarial Contexts. In The Thirteenth International Conference on Learning Representations

  10. [18]

    Zhuohua Li, Maoli Liu, Xiangxiang Dai, and John C.S. Lui. 2025. Towards Efficient Conversational Recommendations: Expected Value of Information Meets Bandit Learning. In Proceedings of the ACM on Web Conference 2025 (Sydney NSW, Australia) (WWW ’25). 4226–4238

  11. [19]

    Zhuohua Li, Maoli Liu, and John C. S. Lui. 2024. FedConPE: Efficient Federated Conversational Bandits with Heterogeneous Clients. In Proceedings of the Thirty- Third International Joint Conference on Artificial Intelligence, IJCAI-24. 4533–4541

  12. [20]

    Manish Raghavan, Aleksandrs Slivkins, Jennifer Wortman Vaughan, and Zhi- wei Steven Wu. 2023. Greedy algorithm almost dominates in smoothed contextual bandits. SIAM J. Comput. 52, 2 (2023), 487–524

  13. [21]

    Manish Raghavan, Aleksandrs Slivkins, Jennifer Vaughan Wortman, and Zhi- wei Steven Wu. 2018. The externalities of exploration and how data diversity helps exploitation. In Conference on Learning Theory . PMLR, 1724–1738

  14. [22]

    Vidyashankar Sivakumar, Steven Wu, and Arindam Banerjee. 2020. Structured lin- ear contextual bandits: A sharp and geometric smoothed analysis. InInternational Conference on Machine Learning . PMLR, 9026–9035

  15. [23]

    Yueming Sun and Yi Zhang. 2018. Conversational recommender system. In The 41st international acm sigir conference on research & development in information retrieval. 235–244

  16. [24]

    Joel A. Tropp. 2011. User-Friendly Tail Bounds for Sums of Random Matrices. Foundations of Computational Mathematics 12, 4 (Aug. 2011), 389–434

  17. [25]

    Zhiyong Wang, Xutong Liu, Shuai Li, and John CS Lui. 2023. Efficient explorative key-term selection strategies for conversational contextual bandits. InProceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 10288–10295

  18. [26]

    Junda Wu, Canzhe Zhao, Tong Yu, Jingyang Li, and Shuai Li. 2021. Clustering of conversational bandits for user preference learning and elicitation. In Pro- ceedings of the 30th ACM International Conference on Information & Knowledge Management. 2129–2139

  19. [27]

    Zhihui Xie, Tong Yu, Canzhe Zhao, and Shuai Li. 2021. Comparison-based conversational recommender system with relative bandit feedback. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1400–1409

  20. [28]

    Shuhua Yang, Hui Yuan, Xiaoying Zhang, Mengdi Wang, Hong Zhang, and Huazheng Wang. 2024. Conversational Dueling Bandits in Generalized Linear Models. arXiv preprint arXiv:2407.18488 (2024)

  21. [29]

    Xiaoying Zhang, Hong Xie, Hang Li, and John CS Lui. 2020. Conversational contextual bandit: Algorithm and application. InProceedings of the web conference

  22. [30]

    Yongfeng Zhang, Xu Chen, Qingyao Ai, Liu Yang, and W Bruce Croft. 2018. Towards conversational search and recommendation: System ask, user respond. In Proceedings of the 27th acm international conference on information and knowledge management. 177–186

  23. [31]

    Canzhe Zhao, Tong Yu, Zhihui Xie, and Shuai Li. 2022. Knowledge-aware conver- sational preference elicitation with bandit feedback. In Proceedings of the ACM Web Conference 2022. 483–492

  24. [32]

    Fixed Interval

    Jinhang Zuo, Songwen Hu, Tong Yu, Shuai Li, Handong Zhao, and Carlee Joe- Wong. 2022. Hierarchical conversational preference elicitation with bandit feed- back. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. 2827–2836. A Appendix...

Pith tools

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