Online Consistency of the Nearest Neighbor Rule

In the realizable online setting, a learner is tasked with making predictions for a stream of instances, where the correct answer is revealed after each prediction. A learning rule is online consistent if its mistake rate eventually vanishes. The nearest neighbor rule (Fix and Hodges, 1951) is a fundamental prediction strategy, but it is only known to be consistent under strong statistical or geometric assumptions: the instances come i.i.d. or the label classes are well-separated. We prove online consistency for all measurable functions in doubling metric spaces under the mild assumption that the instances are generated by a process that is uniformly absolutely continuous with respect to a finite, upper doubling measure.

Paper

References (28)

Scroll for more · 16 remaining

Similar papers

Peer review

Reviewer 4b2W6/10 · confidence 4/52024-06-28

Summary

This paper studies the problem of non-uniform consistency of the nearest neighbor prediction rule when the instances are generated by a random process that has conditional marginals dominated by some underlying reference measure. In particular, it generalizes the well-known results of Cover and Hart (1967) for iid processes and the result of Kulkarni and Posner (1995) for arbitrary processes with strong assumptions on labeling functions. The paper shows that under the uniform dominance condition of the process, the NN rule is consistent for any measurable function, which is further extended to finite sample rates with additional assumptions on both the instance space and processes.

Strengths

I believe studying the consistency of the nearest neighbor prediction rule under a general random process is an important problem for understanding its limitations. This paper makes a significant step toward such a goal. From a philosophical perspective, this work is analogous to the recent line of work in beyond worst-case online learning in the infinite consistency paradigm. The paper also develops several original techniques, such as the control of the nearest neighbor process in Theorem 15, which may be of independent interest. Overall, I think this is an interesting paper and is suitable for publication at NeurIPS.

Weaknesses

My main complaint about this paper is its writing style. The paper leans too much on geometric concepts (such as doubling metric spaces, length spaces, box-counting dimension, and Minkowski content), which, in my opinion, are not necessary to demonstrate the learning-theoretical insights and place an undue burden on the reviewers. Why not just focus on the Euclidean space in the main text and leave the generalizations for the appendix or a journal version? I have also the following specific comments: 1. Can the authors comment on why all the main theorems assume uniform dominance rather than the more relaxed ergodic dominance? My understanding is that the only place where uniform dominance is used is in the Borel-Cantelli argument in the proof of Lemma D.4. Am I right? 2. To be honest, I don't quite understand how the "sequentially-constructed cover tree" works. The sketch from lines 237-253 doesn't really help too much. Are the whole arguments merely to improve the $1/\delta$ to $\log(1/\delta)$? Why can't one simply extend the argument in Example 17 using multiple indicators, if the goal is only consistency? 3. I would recommend the authors move part of Section 7 to the appendix and put more effort in the main text on explaining the proof of Theorem 15. 4. Can the authors comment on how the results from Blanchard (2022) compare to your results? Does the kC1NN rule proposed therein achieve consistency in your setting?

Questions

See above.

Rating

6

Confidence

4

Soundness

3

Presentation

2

Contribution

3

Limitations

N/A

Reviewer XeYp7/10 · confidence 3/52024-07-07

Summary

This paper considers the consistency (or mistake-bound) of the nearest neighbour rule in the realizable online setting when the instances are not necessarily i.i.d. but drawn from a well-behaved stochastic process. The authors prove that when the underlying stochastic process is uniformly dominated, the nearest neighbour rule is consistent for labeling functions that have negligible boundaries. To prove this they introduce the notion of labeling functions on mutually labeling sets, which they then generalize to labeling functions on upper doubling metric measure spaces. Finally they also give convergence rates for smooth stochastic processes, which is a special case of the uniformly dominated setting considered in the paper.

Strengths

The central question of this paper is very well-motivated and an important question in the online learning literature. The paper manages to address the gap of online consistency of the nearest neighbour rule, and outline various interesting conditions for settings in which nearest neighbours rule is consistent. The paper is also quite well written.

Weaknesses

One comment I have regarding the technical novelty is that the proof techniques in this paper are fairly standard across the literature on online learning and consistency of nearest neighbours. I mention this in case the authors would like to highlight something that is not actually a standard technique. This comment does not affect the score. 1. It seems that the paper only proves results for the $1$-NN setting. See point 2 of the questions section. 2. Section 6 seems to serve expository purposes. See point 4 of the questions section.

Questions

1. The authors mention on line 307 that "Dasgupta (2012) ....... considered consistency for non-online settings...." This seems to be a gross oversimplification. I would like a clarification on this point. Dasgupta (2012): Consistency of nearest neighbor classification under selective sampling, COLT 2012. 2. Do the proof techniques extend naturally to the $k$-NN setup or the $k_n$-NN setup? This is not immediately clear to me from the paper. 3. Do the proof techniques extend naturally to the agnostic case where there is no underlying labeling function but instead $Z_n=(X_n,Y_n)$ are jointly drawn according to some underlying stochastic process $Z=(Z_n)_{n>0}$? The proof techniques would then boil down to computing mistake bound against the Bayes optimal classifier in this case. 4. I am not sure about the significance of Theorem 14 or Section 6, and would like a clarification on the following point. The notion of consistency for uniformly dominated stochastic processes should straightforwardly imply a notion of consistency for stochastic processes which are dominated by much weaker notions of convergence. I suspect that Theorem 15 can be generalized to stochastic processes which are mixing, and the parameter $\gamma$ can be replaced by some function of the mixing coefficient.

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

2

Limitations

The paper does a good job at clearly specifying the assumptions used in the setup, but sometimes does not address settings which are not explicitly considered in the paper. Please refer to the Questions section for clarifications.

Reviewer LFbH6/10 · confidence 2/52024-07-13

Summary

This paper studies the nearest neighbor rule in the realizable online setting and closely checks under what assumptions it can achieve online consistency, i.e. the mistake rate eventually vanishes as the number of samples increases. It proves that for all measurable functions in doubling metric spaces when the samples are generated by a uniformly absolutely continuous process with respect to an underlying finite, upper doubling measure, the online consistency is achieved.

Strengths

While the prior work showed the online consistency for nearest neighbor rule under much stronger assumptions (instances are i.i.d. or the label classes are well-separated), this paper improves the understanding of this problem by showing it in the uniformly dominated regime. This result potentially gives us an understanding of other algorithms when studying online learning beyond i.i.d. and smoothed adversary settings.

Weaknesses

While at this point, the review does not notice any significant weakness that needs to be addressed, the writings of the paper could be a bit too technical.

Questions

Why ergodic continuity is related to uniformly absolute continuity such that it is included in the proof?

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the limitations are adequately addressed. There is no potential negative societal impact.

Reviewer 4b2W2024-08-07

I thank the authors for the detailed response. I maintain my current positive rating.

Reviewer XeYp2024-08-08

Thank you for the reply. I have found the replies to be satisfactory, and I will be recommending the paper for acceptance. I will be raising my score to reflect this. As an aside: 1. The $k$-NN setting should be explicitly addressed in the paper. 2. I hope that the authors include a longer discussion section in the appendix of the camera-ready version paper where they can go over whatever results (even trivial) they have for the $k_n$-NN rule with an accompanying note on the hardness of this learning for this setting. 3. The part about Dasgupta (2012) should be explicitly clarified in the paper.

Reviewer LFbH2024-08-13

Official Comment by Reviewer LFbH

Thank you very much for the detailed response and it is very helpful. I would like to keep my current review.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC