Multiclass Transductive Online Learning

We consider the problem of multiclass transductive online learning when the number of labels can be unbounded. Previous works by Ben-David et al. [1997] and Hanneke et al. [2023b] only consider the case of binary and finite label spaces, respectively. The latter work determined that their techniques fail to extend to the case of unbounded label spaces, and they pose the question of characterizing the optimal mistake bound for unbounded label spaces. We answer this question by showing that a new dimension, termed the Level-constrained Littlestone dimension, characterizes online learnability in this setting. Along the way, we show that the trichotomy of possible minimax rates of the expected number of mistakes established by Hanneke et al. [2023b] for finite label spaces in the realizable setting continues to hold even when the label space is unbounded. In particular, if the learner plays for $T \in \mathbb{N}$ rounds, its minimax expected number of mistakes can only grow like $\Theta(T)$, $\Theta(\log T)$, or $\Theta(1)$. To prove this result, we give another combinatorial dimension, termed the Level-constrained Branching dimension, and show that its finiteness characterizes constant minimax expected mistake-bounds. The trichotomy is then determined by a combination of the Level-constrained Littlestone and Branching dimensions. Quantitatively, our upper bounds improve upon existing multiclass upper bounds in Hanneke et al. [2023b] by removing the dependence on the label set size. In doing so, we explicitly construct learning algorithms that can handle extremely large or unbounded label spaces. A key and novel component of our algorithm is a new notion of shattering that exploits the sequential nature of transductive online learning. Finally, we complete our results by proving expected regret bounds in the agnostic setting, extending the result of Hanneke et al. [2023b].

Paper

Similar papers

Peer review

Reviewer Nc668/10 · confidence 4/52024-06-17

Summary

This paper studies transductive online learning in the multiclass setting, where the label space can be unbounded. In the transductive setting, the adversary commits to a sequence of examples and can only adaptively choose labels for a sequence of instances, unlike the pure online setting where the adversary can adaptively choose both the sequence of examples and the labels, or the pure offline where the adversary commits to both a sequence of examples and labels. The main result of this paper extends the results of Hanneke et. al. [2023] by characterizing the optimal mistake bound in the unbounded label case. The proof techniques involve defining and leveraging two new combinatorial parameters termed Level-constrained Branching and Littlestone dimensions. The authors prove that the level-constrained Littlestone dimension characterizes this specific variant of online learning. Hanneke et. al. [2023]: A Trichotomy for Transductive Online Learning.

Strengths

1. This paper characterizes the optimal mistake bound of transductive online learners in the unbounded label space setting in both realizable and agnostic cases. 2. The paper introduces two new combinatorial parameters: the Level-constrained Littlestone dimension $D(C)$ and the Level-constrained Branching dimension $B(C)$, and show that $D(C)$ characterizes transductive online learning in the unbounded label setting. These new parameters are also compared to existing combinatorial parameters like the Natarajan Dimension, DS dimension, and the Graph dimension. 3. The technical parts of the paper involve modifying the Halving technique to work with new notion of shattering using the above newly introduced combinatorial parameters, and a modified version of Littlestone's standard optimal algorithm. Combined with the definition of $D(C)$ and $B(C)$, I feel that these new definitions and techniques could be of independent interest to the learning theory community. 4. This work almost rounds out the literature on regret and mistake bounds in the original transductive online learning setting. Removing the logarithmic factor in the upper bound of Theorem 3 seems to be the final milestone in this setting.

Weaknesses

1. The definitions of expected regret and expected cumulative mistakes in section 2.2 seem to be incomplete. What is the probability distribution that the expectations are taken with respect to? The only clear choice to me is some probability distribution over the label space $\mathcal{Y}$. Since the label space is unbounded (a critical point of this work), the notion of a probability distribution over this unbounded label space should be rigorously defined somewhere in the paper (maybe in the Appendix if there are space constraints). This seems to be an inexplicable oversight in an otherwise highly technical and detailed paper. 2. Even though the regret in the agnostic setting is stated in terms of the level-constrained Littlestone dimension, the proof technique heavily borrows from previous works. The authors state this clearly in the text as well. While the result serves a pedagogical purpose and also lends to completeness of the results, Theorem 4 cannot be counted as a significant contribution. The following weaknesses do not directly impact my scores but should improve the paper's readability. 3. In reference [Brukhim 2022], the DS dimension characterizes multiclass learnability for unbounded label space in the offline setting. It is not immediately clear from the main text why the DS dimension cannot similarly characterize the transductive online setting and why $D(C)$ and $B(C)$ are required in the first place. The answer is possibly buried in Proposition 2, given in Appendix F.2, and should clearly be stated in the main text (even informal statements suffice), given the fact that the new combinatorial parameters are a core contribution of this paper. 4. The proofs in section 3.1 and 3.2 use the standard Halving technique, but make use of the newly defined combinatorial parameters. Both upper bounds share similar overarching ideas with each other, and also with results in previous works for the constrained label space. These two sections can be rewritten in a manner that highlights the power of the Halving technique. While the proofs themselves appear to be correct, if we take the above weakness into consideration, its also not immediately apparent from the main text why the Halving technique cannot be applied using more standard combinatorial parameters such as the DS dimension. Also, I feel that some parts of the proof can be deferred to the appendix (up to the discretion of the authors).

Questions

Minor issues ------------------ 1. It is unclear from the definitions in section 2.2 how optimal regret and optimal mistake bound would be defined differently in the realizable setting. Could the authors expand on this? 2. Equations on lines 210, 285, and 354 should be formatted properly. Future Work Discussions ----------------------------------- Please feel free not to answer these questions if the authors are cramped for time. Not answering these questions will not affect my score. 1. Do the lower bounds of Theorem 3 hold directly for the list transductive online setting? 2. In the online real-valued regression setting, can one reduce a discretized version of the problem to the unbounded label case setting?

Rating

8

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Adequately discussed in the paper.

Reviewer DRE47/10 · confidence 4/52024-07-10

Summary

The paper studies the problem of multiclass transductive online learning where the number of classes can be unbounded. In the transductive setting, the learner receives in advance a sequence of instances $(x_1,…,x_T)$ by an adversary. Then, sequentially at each time step t, it needs to decide a label $\hat{y}_t$, and then the adversary reveals the true label label $y_t$. Given a concept class $C$, in the realizable setting, the adversary must choose the labels according to a concept $c \in C$.

The goal is to design an online algorithm that minimizes the regret, i.e. the total number of mistakes done by the learner compared to the best concept chosen from C in hindsight. The paper fully characterizes the regret for the realizable and the agnostic setting for this problem in the case of an infinite number of classes, extending previous results of Hanneke et al. [2024] limited to a finite number of classes.

Strengths

The paper characterizes the multiclass transductive online learning problem in the case of an infinite number of classes. This is a nice contribution to a fundamental problem. The paper is well written.

Weaknesses

It is not clear whether the tools presented in this paper can be applied to other settings. Specifically, the definitions of level-constrained branching dimension, and level-constrained littlestone dimension seem specific to solve this (still important) problem. (See also question 2 below). In Lines 102-103, it is claimed that finiteness of D(C) and B(C) coincides for |Y| = 2. However, if that’s the case, it seems that Theorem 1 cannot show the Trichotomy of Hanneke et al 2024 for the binary case.

Questions

1. Does the label set Y need to be countable? 2. The DS dimension characterizes the learnability of multiclass learning in the pac setting (Brukhim et al 2022) . Would it be possible to express Theorem 1 using DS and D (i.e., can DS replace the branching dimension, to have a nice parallel between VC and DS?) Typos 207 upper= 211 upper round 359 lowerbounds

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

There is no specific section for limitations. This theory paper is self-contained so I do not believe it is needed.

Reviewer Cx7F7/10 · confidence 4/52024-07-12

Summary

This work continues the study of transductive online classification (a learning setting from the 90s recently reviewed by Hanneke et al. [NeurIPS 2023]. The main result is a trichotomy of possible rates for the general multi-class case (even for the infinite label case) in the realizable setting; answering an open question by Hanneke et al. The three cases are characterized by novel combinatorial dimensions (variant of Littlestone dimension called level-wise Littlestone and ). Additionally they achieve optimal $\tilde{\Theta}(\sqrt{TD(\mathcal{C})})$ (up to log factors) rates in the agnostic setting again determined by the level-wise Littlestone dim. $D(\mathcal{C})$.

Strengths

Timely and interesting paper continuing the recent interest in transductive online classification and related models. Almost tightly characterizes the possible rates for agnostic and realizable learning. --- rebuttal --- changed from 6 to 7

Weaknesses

Not in-depth discussion of previous related work. The gap between $D(C)$ and $B(C)$ in Theorem 3 (in the $B(C)<\infty$ case) is somewhat unsatisfactory. See also questions and limitations below. Minor: * You might want to fix "Hanneke et al. [2024]" to [2023], otherwise you cite a NeurIPS paper in the future. * typo in line 37: should be $c:X\to Y$ not $c:X\mapsto Y$. * typo in line 207: "upper=" * Perhaps hint the additional results (Prop 4, comparison to DS, graph-dim) in the main paper, at least with some short sentences.

Questions

In the agnostic setting for $|Y|=k$ one would expect something like $O(\sqrt{T\mathrm{Ndim}(C)k})$ up to log factors, similar to the bound in the realizable setting in Hanneke et al. 2023 (Theorem B.3). Please relate this to your agnostic bound. Can the gap between $D(C)$ and $B(C)$ be arbitrary? More generally how far can $D(C)$,$B(C)$, and $L(C)$ be from each other. I agree that stating the rate as $O(1)$ makes sense, but it would be nice to have some quantity explicitly giving the rate in this case some Xdim s.t. $O(\mathrm{Xdim}(H))$ (as Ldim does for standard online classification). Note that the tree ranks by Ben-David et al [1997] (see limitations, as well), achieves this specifically also for the worst-sequence/transductive setting (at least for the binary case).

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Please discuss previous work more thoroughly. The tree ranks from the "Online learning versus offline learning" (self-directed, worst-case sequence, etc.) [Ben-David et al. 1997] and similar papers are very much related to the proposed dimensions here (which is only very briefly acknowledged in a short sentence here). E.g., the "level-constrained" variant of the trees in these papers also become the VC-dim for $|Y|=2$. Also similarities to Devulapalli and Hanneke could be discussed more (e.g., the lower bounds there probably apply here too).

Reviewer i2TD6/10 · confidence 3/52024-07-16

Summary

This paper addresses the problem of multiclass transductive online learning with unbounded label spaces. The paper extends previous work on binary and finite label spaces to the more general case of unbounded label spaces. The authors introduce two new combinatorial dimensions - the Level-constrained Littlestone dimension and the Level-constrained Branching dimension - to characterize the optimal mistake bounds in this setting. They establish a trichotomy of possible minimax rates in the realizable setting, showing that the expected number of mistakes can only grow like \theta(T), \theta(log T), or \theta(1).

Strengths

- The paper solves an open problem in online learning theory by characterizing optimal mistake bounds for unbounded label spaces. - The paper is very well written and easy to understand.

Weaknesses

- This paper extends the results of multi class transductive learning to infinite label space but the not clear how important is this setting.

Questions

- Can the authors provide more more intuitive explanation of the Level-constrained Littlestone and Branching dimensions. - Can the authors discuss the computational complexity of their algorithm for this setting?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Reviewer Nc662024-08-08

Thank you for the response and for an excellent paper! I am thoroughly satisfied with the responses and I will be raising my score further in order to stress the contribution of this work.

Reviewer DRE42024-08-11

Thank you for your clarification and your response, and for pointing out to the appendix sections relevant to my questions. (I am sorry for misreading Lines 102-103). After reading the other review and the rebuttal, I still believe this is a good theory paper (and I increased my score accordingly).

Reviewer Cx7F2024-08-12

Thanks for these comments and the remark that in the realizable case the rate is $\Theta(B(C))$ if $T$ is large enough. I raised my score.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC