Pith. sign in

REVIEW 4 major objections 3 minor 42 references

Learning from eXtreme Bandit Feedback

T0 review · 4 major / 3 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Pruning the action space fixes extreme-scale bandit learning

desk verdict A practically motivated paper on debiasing extreme bandit feedback whose central variance-reduction theorem is vacuous under its own assumptions; the empirical method may still work, but the theory needs major repair. read the letter →

arxiv 2009.12947 v2 pith:36WQVPLM submitted 2020-09-27 stat.ML cs.LG

Romain Lopez , Inderjit S. Dhillon , Michael I. Jordan This is my paper · ORCID
classification stat.MLcs.LG MSC 68T0562H30
keywords banditfeedbackimportancesamplingselectiveextrememulti-labelclassificationoff-policylearningRao-Blackwellizationpolicyoptimizationbias-variancetradeoff
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

Batch learning from logged bandit feedback breaks down when the action space has millions of labels: standard importance sampling has impractical variance because the logging policy rarely explores most actions and the reweighting ratios explode. This paper observes that in extreme multi-label classification, most actions are irrelevant for any given example, and proposes a selective importance-sampling (sIS) estimator that reweights only a small per-context subset of actions, namely the top-p actions of the logging policy. The paper proves a bias-variance bound (Theorem 1) showing that when the selected subset covers the truly relevant actions, sIS is unbiased and has strictly lower mean-squared error than full importance sampling. It packages this estimator into a training procedure, POXM, and shows on three public benchmarks (up to 670,000 labels) that POXM systematically improves over the logging policy and over existing bandit-feedback baselines.

What carries the argument

The load-bearing object is the selective importance-sampling estimator (Eq. 7): $$\hat{V}^\Phi_{\mathrm{sIS}}(\pi) = \frac{1}{n}\sum_{i=1}^n \frac{\pi(y_i \mid x_i,\, y_i \in \Phi(x_i))}{\rho(y_i \mid x_i)}\, r_i.$$ It is the Rao-Blackwellization of the full importance-sampling estimator down to the event that the logged action belongs to a per-context selector $\Phi(x)$, which prunes the action space to a tractable subset. Theorem 1 quantifies the tradeoff: bias is bounded by $\Delta$ times $\kappa(\pi,\Psi,\Phi)$, the probability under $\pi$ of landing on relevant actions outside $\Phi$, while the variance reduction appears as a subtracted $\chi^2$-divergence term. In POXM the selector is taken to be the top-$p$ actions of the logging policy, with $p$ chosen from a small grid using the self-normalized importance-sampling value estimate, and the estimator extends to slates via a factored policy with sampling without replacement.

What would settle it

Take a logged bandit dataset with known ground-truth relevant sets, and build a logging policy that assigns low probability to a subset of relevant actions so that some of them fall outside its top-$p$ action set. Compute the sIS estimate of a fixed policy's value and compare its bias to the bound $\Delta\kappa(\pi,\Psi,\Phi)$ from Theorem 1; if the observed bias exceeds the bound, or if POXM trained on this data fails to improve over the logging policy, the paper's central claim would be contradicted.

Watch

Extended reading notes

Core claim

Under a sparse-feedback condition—for every context $x$ there is a small relevant set $\Psi(x)$ whose complement has zero expected reward—and a relaxed overlap condition—the logging policy $\rho$ only needs to place probability at least $\epsilon$ on actions in $\Psi(x)$, not on the whole action space—the paper shows that the $\Phi$-selective importance-sampling estimator $\hat{V}^\Phi_{\mathrm{sIS}}$ enjoys a mean-squared-error bound relative to full importance sampling $\hat{V}_{\mathrm{IS}}$: $$\mathrm{MSE}[\hat{V}^\Phi_{\mathrm{sIS}}(\pi)] \le \mathrm{MSE}[\hat{V}_{\mathrm{IS}}(\pi)] + 2\$\Delta$^2\kappa(\pi,\Psi,\Phi) - \frac{\$sigma^{2}$}{n}\mathbb{E}_x \frac{\$pi^{2}$(\$Phi^{0}$(x)|x)}{\rho(\$Phi^{0}$(x)|x)}$$ where $\kappa$ is the probability that the learned policy puts on relevant actions outside the selector and the subtracted term is a Rao-Blackwellization variance reduction. When the selector $\Phi$ covers $\Psi$, the estimator is unbiased and strictly better. The paper then instantiates $\Phi$ as the logging policy's top-$p$ actions, selects $p$ via self-normalized importance sampling, and builds a policy-gradient procedure, POXM, which it reports to systematically and significantly improve over the logging policy and over prior bandit-feedback methods on three extreme multi-label classification datasets.

Load-bearing premise

The method assumes that for every input there is a small action set carrying all expected reward, and that the logging policy places at least some minimal probability on each action in that set; if a real recommender's top-p list excludes truly relevant actions, the estimator's unbiasedness and variance gains disappear.

Editorial extensions

If this is right

  • Off-policy learning in extreme action spaces becomes feasible without requiring the logging policy to explore the full action space; overlap with a small relevant set per context suffices.
  • The top-$p$ selector with $p$ chosen from the data by self-normalized importance sampling gives a practical bias-variance knob, as illustrated in the paper's logging-policy experiments.
  • POXM systematically improves over the logging policy and over existing bandit-feedback baselines on three public extreme multi-label benchmarks, including one with more than 670,000 labels.
  • Because the estimator only needs the probabilities of the selected actions, the cost of computing the policy's softmax scales with the selector size rather than the total action count.
  • When the selector covers the oracle relevant set, the sIS estimator is unbiased and has lower mean-squared error than full importance sampling, so the gains are statistical as well as computational.

Reading between the lines

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

  • The bias term $\kappa(\pi,\Psi,\Phi)$ suggests a deployable diagnostic: measure how often a logging policy's top-$p$ set covers known relevant actions; a large miss rate bounds the worst-case bias before any training.
  • The same selective reweighting could be composed with doubly robust estimators or learned reward models, potentially shrinking variance further when a predictive reward model is available.
  • The principle generalizes to any logged bandit setting with sparsity in action-value structure—slate recommendation, large retrieval, or search—where only a small candidate set per context is plausible.
  • The paper's experiments with more stochastic logging policies suggest the improvement may degrade as top-$p$ coverage drops; a systematic sweep of exploration temperatures would map where POXM stops beating the logging policy.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

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 batch learning from bandit feedback in extreme-scale action spaces, as arising in XMC-style recommendation problems. It introduces a selective importance sampling (sIS) estimator that, instead of importance-weighting all actions, restricts the estimator to a per-instance action selector Φ, motivated by the assumption that most actions have minimal expected reward. Theorem 1 states a bias-variance tradeoff for sIS relative to ordinary importance sampling, with a variance-reduction term proportional to a quantity σ² and a bias term controlled by the policy mass on Ψ∩Φ0. The paper then proposes POXM, which plugs sIS into a slate-policy objective with a renormalized policy πΦ and a BanditNet-style reward translation, and evaluates POXM on three XMC datasets converted to bandit feedback. The reported results claim that POXM systematically and significantly improves over the logging policy, the direct method, BanditNet, and a partial-matching variant.

Significance. The problem addressed is timely and practically important: off-policy learning from logged feedback with millions of actions is central to recommendation systems, and the paper is, to my knowledge, among the first to bring BLBF methodology to XMC-scale action spaces. The supervised-to-bandit conversion and the use of AttentionXML as a base model are sensible components of an experimental pipeline, and the idea of pruning the action selector to reduce importance-sampling variance is appealing. However, the paper's central theoretical claim is not supported: under the paper's own Assumption 2, the alleged variance-reduction term vanishes, and the proof of the main theorem contains additional technical errors. The experimental section also lacks confidence intervals and significance tests, so the word 'significantly' in the abstract is not justified. The practical method may still have value, but the manuscript as written does not establish its main claims.

major comments (4)
  1. [§3, Theorem 1 and Appendix proof] The claimed variance advantage of sIS is not supported under the paper's own Assumption 2. After the 'without loss of generality' normalization ∇=0 in §3, rewards satisfy r∈[0,Δ] and E[r|x,y]=0 for every y∈Ψ0(x); nonnegativity then forces r=0 almost surely on Ψ0. Consequently, for any selector Φ with Ψ⊆Φ, the sIS estimator (Eq. 7) and the ordinary IS estimator (Eq. 1) have identical sample contributions, since terms with y∈Φ0 contribute zero reward in both estimators. The variance-reduction term in Eq. (10) vanishes because σ²=inf E[r²|x,y]=0, so Eq. (9) reduces to an inequality with no strict improvement (equality when κ=0). The theorem therefore does not establish the 'significantly more favorable bias-variance regime' claimed in the abstract; sIS as analyzed is a truncation, not a Rao-Blackwellization that conditions on a sufficient statistic.
  2. [Appendix, Eqs. (21)–(25)] The second-moment calculation in the proof is incorrect as written. Eq. (21) contains the conditional second moment E_{P(r|x,y)}[r²], but Eq. (22) replaces this quantity with v(x,y)=E[r|x,y]; this substitution is not valid in general, and it is inconsistent with the theorem statement, which defines σ²=inf E[r²|x,y]. Furthermore, the proof asserts that the chi-square divergence in Eq. (24) is 'greater or equal to 1.' This is false for the standard chi-square divergence, which is nonnegative and equals 0 when the two distributions coincide. Consequently, the lower bound in Eq. (25), and hence the variance-reduction term in Eq. (10), is not established.
  3. [§4, Eqs. (11)–(13) versus Eq. (7)] The POXM objective is not the sIS estimator analyzed in Theorem 1. In Eq. (11), the importance weight uses πΦ, the policy restricted to Φ, and Eq. (12) renormalizes the policy over Φ(x); Eq. (13) additionally applies a BanditNet-style reward translation. The experimental gains in Table 2 could therefore come from the pruning/renormalization of the learned policy or from the reward translation, rather than from the variance-reduction property of sIS proved in Theorem 1. No ablation is provided that would separate these mechanisms, so the paper's central attribution of POXM's empirical success to the sIS estimator is unsupported.
  4. [§5, Table 2 and Results] The experimental section reports a single run for each method and provides no confidence intervals, standard errors, or significance tests. The abstract's statement that POXM 'systematically and significantly improves over all baselines' is therefore not backed by the evidence presented. This concern is especially acute on Amazon-670K, where the absolute gaps between POXM and the closest baseline are modest (e.g., R@5 of 23.72 versus 20.14), and run-to-run variability could be material.
minor comments (3)
  1. [§3, Theorem 1 statement] The notation R^d×[K] in the definition of σ² is undefined; the paper elsewhere uses L for the number of actions, so either K should be replaced by L or defined explicitly.
  2. [Appendix proof of Theorem 1] The proof switches inconsistently between 'sIS' and 'sIPS' (e.g., Eq. (15) and surrounding text) and between 'IPS' and 'sIPS' in Eq. (20); the notation should be harmonized throughout.
  3. [Appendix, Table 4] The wPOXM results are reported only as relative percentage changes without absolute values or error bars, making the claimed +4.77% improvement in PSR@3 difficult to evaluate relative to run-to-run variability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the theoretical variance-reduction overclaim is a correctness issue, not a derivation from fitted targets.

full rationale

No load-bearing step in the paper reduces to its own inputs. The sIS estimator (Eq. 7) is defined independently of Theorem 1, and Theorem 1 is derived from the stated Assumptions 2 and 3 rather than fitted to the experimental results. The selector size p is chosen on training data via the SNIS estimator, and the translation parameter lambda is also selected on a training-data grid as in BanditNet; neither is calibrated to test-set metrics. The experimental evaluation is an external supervised-to-bandit conversion on standard XMC datasets against named baselines (BanditNet, direct method, partial-matching variants), so the empirical claims are self-contained and falsifiable. The main caveat is a mathematical overclaim rather than circularity: after the 'w.l.o.g.' normalization abla = 0, Assumption 2 makes the irrelevant-action reward nonnegative with zero expectation, hence almost surely zero; consequently, when Ψ ⊆ Φ the sIS sum in Eq. (7) and the IS sum in Eq. (1) are termwise identical, and the variance-reduction term in Theorem 1 vanishes because sigma^2 = 0. Thus the paper's assertion that sIS has 'more favorable performance' when Φ covers Ψ is not actually established by Theorem 1. This is a rigor/correctness concern outside the circularity axis.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The core estimator rests on the sparse-feedback and overlap assumptions (Assumptions 2 and 3); without them the top-p estimator has uncontrolled bias. The empirical claims additionally depend on the logging policy being built from a supervised model trained on a fraction of the ground-truth labels, which is favorable to the top-p construction.

free parameters (5)
  • p (number of top logging-policy actions kept by the selector Phi_p) = grid {10,20,50,100}, selected per dataset via SNIS
    Controls the bias-variance tradeoff in Theorem 1; larger p covers more of Psi but adds variance.
  • lambda (reward translation in the POXM objective) = grid {0.7,0.8,0.9,1.0}, selected via SNIS
    BanditNet-style shift to reduce propensity overfitting, Eq. (13).
  • slate size ell = 5
    Fixed for all experiments; the action-space cardinality of the slate in Eqs. (3)-(5).
  • logging policy training fraction alpha = 0.2 (EUR-Lex, Amazon-670K), 0.1 (Wiki10-31K)
    Controls how much of the ground-truth labels are used to build the logging policy, which determines how relevant top-p actions are.
  • logging temperature T and Gumbel scale beta = per-dataset values in Table 3 (e.g., T=2, beta=0 for EUR-Lex A)
    Set by hand to control the stochasticity of the logging policy; used to create robustness scenarios on EUR-Lex.
assumptions (6)
  • domain assumption Sparse feedback condition (Assumption 2): for every x, [L] has a subset Psi(x) such that all actions outside Psi(x) have expected reward Delta=0.
    Central modeling premise enabling the restrictive estimator; if false, the sIS estimator has unbounded bias.
  • domain assumption Psi-overlap condition (Assumption 3): logging policy rho(y|x) > epsilon for all y in Psi(x).
    Ensures logged data covers relevant actions; if false, importance weights are infinite or zero and the estimator fails.
  • domain assumption Phi(x) subset supp rho(.|x) almost surely for the chosen selector.
    Required for finite importance weights in Eq. (7); assumed when constructing top-p selectors.
  • domain assumption Reward decomposes as a sum of independent per-action contributions, P(R|x,Y)=prod_j P(r_j|x,y_j).
    Used to write V(pi) as a sum over slate positions in Eq. (5); excludes rank-dependent or interaction rewards.
  • domain assumption Policy factorizes autoregressively over slate positions, pi(Y|x)=prod_j pi(y_j|x,y_1:j-1).
    Used to define per-position importance weights; the authors note this is without loss of generality when action order is logged.
  • standard math Sklar's theorem on copulas.
    Invoked to justify deriving joint slate distributions from marginals; standard result.
invented entities (1)
  • Psi(x), the oracle action selector
    purpose: Formalizes the unknown set of relevant actions for each context, enabling the selective importance sampling decomposition.
    Introduced as a theoretical construct; no direct observable is provided, and the paper only postulates its existence through Assumption 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning from eXtreme Bandit Feedback." pith.science (2026). https://pith.science/paper/36WQVPLM

@misc{pith2026200912947,
  author       = {Pith},
  title        = {Pith review of: Learning from eXtreme Bandit Feedback},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/36WQVPLM}},
  note         = {Machine review of arXiv:2009.12947}
}
read the original abstract

We study the problem of batch learning from bandit feedback in the setting of extremely large action spaces. Learning from extreme bandit feedback is ubiquitous in recommendation systems, in which billions of decisions are made over sets consisting of millions of choices in a single day, yielding massive observational data. In these large-scale real-world applications, supervised learning frameworks such as eXtreme Multi-label Classification (XMC) are widely used despite the fact that they incur significant biases due to the mismatch between bandit feedback and supervised labels. Such biases can be mitigated by importance sampling techniques, but these techniques suffer from impractical variance when dealing with a large number of actions. In this paper, we introduce a selective importance sampling estimator (sIS) that operates in a significantly more favorable bias-variance regime. The sIS estimator is obtained by performing importance sampling on the conditional expectation of the reward with respect to a small subset of actions for each instance (a form of Rao-Blackwellization). We employ this estimator in a novel algorithmic procedure -- named Policy Optimization for eXtreme Models (POXM) -- for learning from bandit feedback on XMC tasks. In POXM, the selected actions for the sIS estimator are the top-p actions of the logging policy, where p is adjusted from the data and is significantly smaller than the size of the action space. We use a supervised-to-bandit conversion on three XMC datasets to benchmark our POXM method against three competing methods: BanditNet, a previously applied partial matching pruning strategy, and a supervised learning baseline. Whereas BanditNet sometimes improves marginally over the logging policy, our experiments show that POXM systematically and significantly improves over all baselines.

Figures

Figures reproduced from arXiv: 2009.12947 by the authors.

Figure 1
Figure 1. Expected R@5 and CDF of the logging policy for the top-k action for each XMC dataset. Exploration is limited to a subset of relevant actions. derive all joint distributions starting from the corresponding one-dimensional marginals (Sklar 1959). In this work, we focus on the case of ordered sampling without replacement to respect an important design restriction: the slate Y must not have redundant actions. For the j-… view at source ↗
Figure 2
Figure 2. Data-driven selection of p on the EUR-LeX dataset. Left: logging policy statistics for three randomization scenarios (A, B, C, described in appendix). Middle: R@5 performance for each POXM variant and each logging policy. Right: SNIS estimates used for selection of p in POXM. LeX but not for the other datasets. For all choices of logging policy in [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    Achiam, J.; Held, D.; Tamar, A.; and Abbeel, P. 2017. Con- strained Policy Optimization. In International Conference on Machine Learning

    work page 2017

  2. [2]

    Agrawal, R.; Gupta, A.; Prabhu, Y .; and Varma, M. 2013. Multi-label learning with millions of labels: Recommending advertiser bid phrases for web pages. InInternational World Wide Web Conference

    work page 2013

  3. [3]

    R.; and Mihaela Van Der Schaar

    Atan, O.; Zame, W. R.; and Mihaela Van Der Schaar. 2018. Counterfactual Policy Optimization Using Domain- Adversarial Neural Networks. In ICML CausalML Work- shop

    work page 2018

  4. [4]

    Babbar, R.; and Sch ¨olkopf, B. 2017. Dismec: Distributed sparse machines for extreme multi-label classification. In International Conference on Web Search and Data Mining

    work page 2017

  5. [5]

    Babbar, R.; and Sch ¨olkopf, B. 2019. Data scarcity, robust- ness and extreme multi-label classification. Machine Learn- ing

    work page 2019

  6. [6]

    Bai, X.; Guan, J.; and Wang, H. 2019. A Model-Based Re- inforcement Learning with Adversarial Training for Online Recommendation. In Advances in Neural Information Pro- cessing Systems

    work page 2019

  7. [7]

    Bhatia, K.; Dahiya, K.; Jain, H.; Mittal, A.; Prabhu, Y .; and Varma, M. 2016. The extreme classification repository: Multi-label datasets and code. URL http://manikvarma.org/ downloads/XC/XMLRepository.html

    work page 2016

  8. [8]

    Bhatia, K.; Jain, H.; Kar, P.; Varma, M.; and Jain, P. 2015. Sparse local embeddings for extreme multi-label classifica- tion. In Advances in Neural Information Processing Sys- tems

    work page 2015

Show all 42 references
  1. [9]

    Casella, G.; and Robert, C. P. 1996. Rao-Blackwellisation of sampling schemes. Biometrika

    1996

  2. [10]

    Chang, W.-C.; Yu, H.-F.; Zhong, K.; Yang, Y .; and Dhillon, I. 2019. X-BERT: eXtreme Multi-label Text Classifica- tion with using Bidirectional Encoder Representations from Transformers. arXiv

    2019

  3. [11]

    Degris, T.; White, M.; and Sutton, R. S. 2012. Off-policy actor-critic. In International Conference on Machine Learn- ing

    2012

  4. [12]

    Dudik, M.; Langford, J.; and Li, L. 2011. Doubly Robust Policy Evaluation and Learning. In International Confer- ence on Machine Learning

    2011

  5. [13]

    Gentile, C.; and Orabona, F. 2014. On multilabel classifi- cation and ranking with bandit feedback. The Journal of Machine Learning Research

    2014

  6. [14]

    Reddi, S.; and Kumar, S. 2019. Breaking the Glass Ceiling for Embedding-Based Classifiers for Large Output Spaces. In Advances in Neural Information Processing Systems

    2019

  7. [15]

    Jain, H.; Prabhu, Y .; and Varma, M. 2016. Extreme multi- label loss functions for recommendation, tagging, ranking & other missing label applications. In International Confer- ence on Knowledge Discovery and Data Mining

    2016

  8. [16]

    Pfannschmidt, K.; Klerx, T.; and Hullermeier, E. 2016. Extreme F-measure maximization using sparse probabil- ity estimates. In International Conference on Machine Learning

    2016

  9. [17]

    Joachims, T.; Swaminathan, A.; and de Rijke, M. 2018. Deep Learning with Logged Bandit Feedback. In Interna- tional Conference on Learning Representations

    2018

  10. [18]

    Joachims, T.; Swaminathan, A.; and Schnabel, T. 2017. Un- biased learning-to-rank with biased feedback. In Interna- tional Conference on Web Search and Data Mining

    2017

  11. [19]

    Johansson, F.; Shalit, U.; and Sontag, D. 2016. Learning Representations for Counterfactual Inference. In Interna- tional Conference on Machine Learning

    2016

  12. [20]

    Khandagale, S.; Xiao, H.; and Babbar, R. 2020. Bonsai- diverse and shallow trees for extreme multi-label classifica- tion. Machine Learning

    2020

  13. [21]

    Langford, J.; Strehl, A.; and Wortman, J. 2008. Exploration scavenging. In International Conference on Machine learn- ing

    2008

  14. [22]

    Lefortier, D.; Swaminathan, A.; Gu, X.; Joachims, T.; and de Rijke, M. 2016. Large-scale Validation of Counterfac- tual Learning Methods: A Test-Bed. In What If workshop: NeurIPS

    2016

  15. [23]

    Y .; and Zitouni, I

    Li, L.; Kim, J. Y .; and Zitouni, I. 2015. Toward predicting the outcome of an A/B experiment for search relevance. In International Conference on Web Search and Data Mining

    2015

  16. [24]

    Vinay, V .; and Wen, Z. 2018. Offline evaluation of ranking policies with click models. In International Conference on Knowledge Discovery and Data Mining

    2018

  17. [25]

    Liu, J.; Chang, W.-C.; Wu, Y .; and Yang, Y . 2017. Deep learning for extreme multi-label text classification. In Inter- national Conference on Research and Development in Infor- mation Retrieval

    2017

  18. [26]

    Lopez, R.; Li, C.; Yan, X.; Xiong, J.; Jordan, M.; Qi, Y .; and Song, L. 2020. Cost-Effective Incentive Allocation via Structured Counterfactual Inference. In AAAI Conference in Artificial Intelligence

    2020

  19. [27]

    McAuley, J.; and Leskovec, J. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. In International Conference on Recommender Systems

    2013

  20. [28]

    L.; and F ¨urnkranz, J

    Mencia, E. L.; and F ¨urnkranz, J. 2008. Efficient pairwise multilabel classification for large-scale problems in the legal domain. In Joint European Conference on Machine Learn- ing and Knowledge Discovery in Databases

    2008

  21. [29]

    Dean, J. 2013. Distributed representations of words and phrases and their compositionality. In Advances in neural information processing systems

    2013

  22. [30]

    Prabhu, Y .; Kag, A.; Harsola, S.; Agrawal, R.; and Varma, M. 2018. Parabel: Partitioned label trees for extreme classi- fication with application to dynamic search advertising. In International World Wide Web Conference

    2018

  23. [31]

    Prabhu, Y .; Kusupati, A.; Gupta, N.; and Varma, M. 2020. Extreme Regression for Dynamic Search Advertising. In International Conference on Web Search and Data Mining. Rahul; Dahiya, H.; and Singh, D. 2019. A Review of Trends and Techniques in Recommender Systems. In Internati...

    2020

  24. [32]

    Sklar, A. 1959. Fonctions de repartition a n-dimensions et leurs marges. Publications de l’Institut Statistique de l’Universit´e de Paris

    1959

  25. [33]

    Su, Y .; Wang, L.; Santacatterina, M.; and Joachims, T. 2019. CAB: Continuous adaptive blending for policy evaluation and learning. In International Conference on Machine Learning

    2019

  26. [34]

    Swaminathan, A.; Krishnamurthy, A.; Agarwal, A.; Dudik, M.; Langford, J.; Jose, D.; and Zitouni, I. 2017. Off-policy evaluation for slate recommendation. In Advances in Neural Information Processing Systems

    2017

  27. [35]

    Tagami, Y . 2017. AnnexML: Approximate nearest neigh- bor search for extreme multi-label classification. In Interna- tional Conference on Knowledge Discovery and Data Min- ing

    2017

  28. [36]

    Wang, L.; Bai, Y .; Bhalla, A.; and Joachims, T. 2019. Batch Learning from Bandit Feedback through Bias Corrected Re- ward Imputation. In ICML Workshop on Real-World Se- quential Decision Making

    2019

  29. [37]

    Wang, Y .; Yin, D.; Jie, L.; Wang, P.; Yamada, M.; Chang, Y .; and Mei, Q. 2016. Beyond ranking: Optimizing whole-page presentation. In International Conference on Web Search and Data Mining

    2016

  30. [38]

    Wu, H.; and Wang, M. 2018. Variance Regularized Counter- factual Risk Minimization via Variational Divergence Min- imization. In International Conference on Machine Learn- ing

    2018

  31. [39]

    Wydmuch, M.; Jasinska, K.; Kuznetsov, M.; Busa-Fekete, R.; and Dembczynski, K. 2018. A no-regret generalization of hierarchical softmax to extreme multi-label classification. In Advances in Neural Information Processing Systems

    2018

  32. [40]

    E.-H.; Huang, X.; Dai, W.; Ravikumar, P.; Dhillon, I.; and Xing, E

    Yen, I. E.-H.; Huang, X.; Dai, W.; Ravikumar, P.; Dhillon, I.; and Xing, E. 2017. Ppdsparse: A parallel primal-dual sparse method for extreme classification. In International Conference on Knowledge Discovery and Data Mining

    2017

  33. [41]

    Dhillon, I. 2016. Pd-sparse: A primal and dual sparse ap- proach to extreme multiclass and multilabel classification. In International Conference on Machine Learning

    2016

  34. [42]

    Zubiaga, A. 2012. Enhancing navigation on Wikipedia with social tags. arXiv . Proofs Theorem 1 (Bias-variance tradeoff of selective importance sampling). Let R andρ satisfy Assumptions 2 and 3. Let Φ be an action selector such that Φ(x) ⊂suppρ(⋅/divides.alt0x) almost surely in...

    2012

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.