NeurIPS 2020

Estimation of Skill Distribution from a Tournament

Review 1

Summary and Contributions: - Theoretical: main contribution. The authors show that it is possible, with the proposed algorithm, to estimate a density of a latent variable (skill) from pairwise comparisons whose discrete (win/loss) outcomes are governed by a BTL probability model whose parameters correspond to the latent variables, matching in a nearly-optimal way (up to poly-logarithmic terms) known minimax MSE lower bounds for standard direct density estimation. Significance: High. Other contributions: new minimax lower bound in terms of l_0 and l_inf losses for the skill parameters, showing that the learning rate of the rank centrality algorithm (proposed in prior work and used as a component in the algorithm proposed here) is near-optimal (up to poly-logarithmic terms). - Algorithmic: The algorithm is key to prove the minimax MSE upper bound for the indirect estimation problem. Essentially, it combines two existing algorithms, namely: rank centrality to estimate skills of a BTL model from pairwise comparisons and Parzen-Rosenblatt kernel density estimation. The algorithm is computationally tractable but requires some parameters that in practice are not known, and thus, requires some ad-hoc steps to use it in practice. Significance: Medium. - Methodological: This paper proposes to use the entropy of the latent skill distribution to quantify the role of chance in tournaments, therefore providing a rigourous statistical approach to this problem of importance for gambling regulations. Significance: Medium. In summary, this is not a paper proposing an algorithm with guarantees but more an algorithm that allows to prove an upper bound on the MSE for the indirect density estimation problem. In addition, it proposes the elegant idea of quantifying the role of chance in games by using the entropy of the estimated density.

Strengths: This is a strong theoretical paper with rigorous proofs and proper positioning with respect to prior work. The theoretical results are novel and very significant since they provide an achievable bound for the indirect density estimation problem nearly-matching the direct estimation learning optimal minimax rate. The approach of using the negative entropy of the skill distribution for quantifying the role of chance in tournaments is novel and elegant, while previous work use ad-hoc measures. I think that the work is highly relevant to the NeurIPS community working on Ranking Learning, because the algorithm and the derived bounds constitute an important reference for future work.

Weaknesses: The limitations of this work are essentially along the empirical evaluation axis. The case where the number of comparisons k is variable would deserve a better treatment since it is usually the case as it is shown in the experiments. In practice, the algorithm can't be applied as proposed because the bandwitdh computation requires unknown parameters and therefore, in the experiments, the authors resort to ad-hoc practices that are not detailed. Moreover, in the experiments, an ad-hoc Laplace smoothing is performed to regularize in the small data regime making. The experimental results are illustrative but it is not clear how these ad-hoc decisions impact the results. --- EDIT after response Thank you for your feedback. Your clarification on how to handle the case of variable number of comparisons should help the reader. Regarding the ad-hoc steps, I did see the precise values (note that for the reproducibility question I answered "Yes"). I was referring to how you got these values which is answered by the second part of your answer.

Correctness: I checked high level arguments of the main text and some of the proofs in the Appendix and I didn't find any errors.

Clarity: This is a very well written paper that I enjoyed reading. The math is very well detailed and explained in the Appendix but also properly summarized in the main text. There is also a very good "Related work" section in the appendix that would deserve to be in the main paper if space constraints allowed it.

Relation to Prior Work: It properly positions it in the context of prior work making easy to understand the relevance of the results.

Reproducibility: Yes

Additional Feedback: The case where the number of comparisons k is variable would deserve a better treatment since it is usually the case as it is shown in the experiments. In particular, I suggest to clarify what kind of alteration of the bandwidth should be done when k is variable (Line 151). In general, it would be helpful to explain how the ad-hoc steps are performed since no code is provided. As a suggestion for future work, a possible way of avoiding the ad-hoc steps would be to replace the PR kernel density estimation by a Bayesian or MDL principled approach like in Rissanen, Jorma, Terry P. Speed, and Bin Yu. "Density estimation by stochastic complexity." IEEE Transactions on Information Theory 38.2 (1992): 315-323. In the Mutual Funds experiment, the performance could be considered as a direct observation of skill. Therefore it would be interesting to compare a direct density estimation vs the one based on the (somewhat artificial) pairwise comparisons. Minor comments: By putting the different plot distributions in the same plot, one could be tempted to compare them directly which does not make sense since the skill estimated from different graphs are not comparable. It would be helpful for the reader to clarify this. Line 296: "other" appears twice

Review 2

Summary and Contributions: This paper addresses the problem of learning the probability density function of Bradley-Terry skill parameters given pairwise comparison outcomes. The main contribution of the paper is an algorithm that estimates the PDF of skill parameters with near-optimal minimax error rate. This rate is shown to be no worse than the rate obtained using the skill parameters themselves (i.e., directly using samples of the PDF). Informally, this results shows that one does not "lose much" by having access to skill parameters only through noisy comparison outcomes.

Strengths: To the best of my knowledge, this is the first paper that addresses the question of estimating the distribution from which Bradley-Terry model parameters are sampled (most prior work focuses on reliably estimating the parameters themselves). It builds on and extends a long line of work published at NeurIPS and other machine learning venues. The results presented in this paper are non-trivial and quite elegant. The main take-away—that, in some sense, the skill disitribution can be estimated as efficiently from comparison outcomes as it can be from samples directly—is surprising, and will certainly be of interest to preference learning researchers & practitioners. In passing, the authors, also prove a few bounds on the error of the Rank Centrality, results which are interesting contributions by themselves.

Weaknesses: My main concern / question is about the interpretation of the resulting skill distributions in Section 4. Bradley-Terry models are typically parametrized in one of two different ways: 1. strictly positive parameters, such that the ratio of the parameters corresponding to two items equals the odds of one item winning over the other (let us call these parameters "skills") 2. real-valued parameters, such that the probability of one item winning over the other is equal to the logistic function of the difference in parameters (let us call these parameters "logits") In this paper, the authors work with the parametrization (1). I would argue that comparing entropies of distributions in that parametrization leads to results that are somewhat difficult to interpret. In particular, it is not clear to me what a uniform skill distribution over [eps, 1] means in terms of expected outcomes. Indeed, the comparison outcome probabilities between teams with skill eps vs. 2*eps are comparable to those between teams with skills 0.5 and 1. But a uniform distribution puts much more mass on the interval (0.5, 1) as it does on the interval (eps, 2*eps). In my opinion, Figures 1c and 1f illustrate this: "World" and "English" seem to have very different skill distributions (in terms of "fan experience"), yet result in a similar differential entropy scores. I suspect it is much easier to interpret distributions over logits. Do the authors have any comments on this? --- EDIT after response Thank you for your feedback. In terms of applications / interpretation, I am still not convinced: if there are n >> 2 teams - a uniform distribution over logits has a fairly intuitive interpretation. In expectation, prob(i wins over j) depends only on the rank difference between teams i and j - Whereas a uniform distribution over skills is hard to make sense of. In expectation, match outcomes will be much "noisier" / uncertain between teams at the top of the ranking, much less so for teams at the bottom of the ranking. As such, I believe the link between skill distribution entropy and sports intuition & fan experience could still be clarified. Nevertheless, since the rest of the results developed in the paper are rigorous, strong & interesting, I continue to be very much in favor of accepting this paper.

Correctness: I did not check the proofs systematically. The main arguments behind the various results appear to be plausible.

Clarity: The paper is very well written and was a pleasure to read.

Relation to Prior Work: Even though the related work section is deferred to the supplementary material, the authors do a good job at positioning their work in the context of related literature "along the way", as they introduce the problem and develop their method.

Reproducibility: Yes

Additional Feedback:

Review 3

Summary and Contributions: This paper regards learning the skill distribution of a population of agents from pairwise games under the assumption that outcome likelihoods are governed by a multinomial logit model. At a high level, the paper’s algorithm first estimates the skill parameters of each agent, then uses these estimate parameters to estimate the distribution from which they were sampled. The paper suggests that the entropy of the distribution provides a quantitative measure of skill/competitiveness/excitingness. To support this claim, the paper analyzes data from cricket, soccer, and mutual funds.

Strengths: Overall, the reviewer feels that the paper is well-written and does a good job at explaining the important ideas in language understandable to those without a statistical background. The paper’s goals, claimed contributions, and experiments are all clear.

Weaknesses: The reviewer has minor criticisms regarding the paper’s use of the language skill, excitingness, and competitiveness. Calling negative the negative entropy of a skill distribution the “overall skill score” reads as if it is a measurement of the skill of the participants, whereas it is meant to be read as a measurement of the variation of the skill of the participants. Suggesting that this corresponds to excitingness also seems not quite right. Many pundits feel that recent NBA seasons have been exciting, despite that there were massive skill discrepancies between the top few teams and the rest of the league. It may also be the case that a league of closely matched teams has many one-sided/non-competitive games, or that a league of not closely matched teams has many close games.

Correctness: The reviewer does not possess the statistical prowess to follow the paper’s theoretical results.

Clarity: Yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: The reviewer has one significant question of the paper regarding the application of its method as a quantitative measurement. Why is the entropy of the estimated skill distribution from which the samples were drawn preferred to the entropy of the estimated empirical distribution? The reviewer’s intuition is that the latter would better quantify the skill/competitiveness/excitingness of the games. If this intuition is correct, what reason is there to estimate to estimate the underlying skill distribution? The reviewer’s opinion of this paper depends on the answer to this question. The reviewer is very willing to raise the assessed score accordingly. AFTER REBUTTAL: The reviewer's main concern with the submission was the motivation behind estimating the skill PDF. Based on the rebuttal and the other reviews, it seems that this motivation is well-founded. The reviewer increased the assessed score accordingly.

Review 4

Summary and Contributions: This paper proposes an algorithm to learn BTL model. Different from the usual setting, this paper assumes that the parameter of each item (team, fund manager, etc.) follows a distribution and the proposed algorithm aims at estimating this distribution. This distribution gives information on how different the items are. The algorithm has two steps: estimating the parameter for each item, and estimating the distribution using the estimated parameter in the first step. The authors provide theoretical bound on the accuracy of the algorithm and some experiments that validate their motivation.

Strengths: The main contribution of this paper is theoretical. Theorems 1 and 2 aims at characterizing the accuracy of learning the parameter of BTL and Theorem 3 characterizes the accuracy of estimating the overall distribution. I did not check the proofs, but these theorems seem nontrivial.

Weaknesses: 1. This work is not well-motivated. It seems estimating the distribution is not necessary since the variance of the learned parameters for each item (teams, fund managers, etc.) already indicates whether the parameters of all items closely center somewhere or spread uniformly. So I don't see the point of a sophisticated second step in the algorithm. This paper does not provide a comparison in the experiments between their approach and the naive approach mentioned above either. 2. Most of this paper seems very technical but the theorems do not qualify as theorems. "sufficiently large constants" and "sufficiently large n" appear in every theorem statement but the authors fail to specify how large is sufficient. These are necessary for theorems.

Correctness: I don't see obvious errors, but the statements of theorems need to be rewritten.

Clarity: Yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: I appreciate the authors' response and have updated my score accordingly.