acceptodds
Under review as a conference paper at ICLR 2027

The Unactionability of Concept Drift Detection

Abstract

Change detectors are commonly used alongside fixed heuristic rules as adaptation mechanisms in online learning algorithms deployed in non-stationary environments. This work investigates whether a change detector signal contains sufficient information to determine an appropriate model adaptation strategy. We show theoretically that a detector can be unactionable: distinct post-drift distributions may produce the same detector signal while having different optimal adaptation strategies. We further derive an information-theoretic bound showing that the information available to an adaptation policy is limited by the information exposed by the detector, and that this can lead to unavoidable misadaptation when the entropy of the optimal adaptation decision exceeds detector capacity. We evaluate these results on controlled synthetic streams and study their practical implications across 24 streams from the USP DS repository. Empirical results show substantial variation in the utility of fixed detector-triggered adaptation across stream-learning architectures and datasets, motivating adaptation policies that utilise more information than a fixed detector signal alone.

Primary area:
learning theory
Keywords:
concept drift, data streams, learning theory, online learning, evolutionary data, non-stationarity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_ZTiYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies whether a change detector signal is sufficient to determine an appropriate model adaptation strategy. It introduces unactionability through post-drift distributions that produce the same detector signal but favour different adaptation operators. It also derives an information-theoretic bound relating detector capacity to the probability of misadaptation. Synthetic experiments illustrate the proposed drift scenarios, while a case study across 24 streams from the USP DS repository examines the utility of fixed detector-triggered adaptation across different learning architectures.

Critical strengths
  1. The paper addresses an important distinction between change detection and model adaptation. Detecting a change does not by itself determine whether to retain, reset, or partially adapt a learner. The finite-horizon risk formulation is well motivated, as it accounts for both the immediate cost of an adaptation operation and the learner’s subsequent recovery.
  2. The case study provides useful evidence on the effects of fixed adaptation policies. The evaluation covers decision trees, ensembles, kNNs, and neural networks, and includes comparisons with and without change detectors. The results show that detector-triggered adaptation can help some architectures and hurt others. This supports the paper’s argument that adaptation policies should account for the learner and the data stream.
Critical weaknesses
  1. The proofs of Theorems 2–4 do not establish the claimed finite-horizon risk comparisons. In Section 3.1, the learner continues to update after either the No-Action or Reset operator is applied. However, the proof sketches in Appendices A.2–A.4 assume that retaining the existing learner leads to persistent error, while resetting allows it to learn the new decision boundary. The retained learner can also learn from post-drift samples. Without specifying the online learning rule, reset initialization, and horizon (H), it is unclear why reset should have strictly lower finite-horizon risk. Appendix A.1 also contains a smaller proof error: the same lower bound does not necessarily hold after randomizing over adaptation operators. For example, excess-risk vectors ((0,1)) and ((1,0)) have worst-case risk 1 individually, but their uniform mixture has worst-case risk (1/2).
  2. The information bottleneck analysis applies to a restricted adaptation policy. Theorem 5 assumes that the adaptation policy is conditioned strictly on the detector signal. In practice, a policy may also use the learner state, previous detector signals, and the time of detection. The one-bit bound therefore applies to a single binary signal, rather than necessarily to all information available for adaptation. Moreover, a non-zero probability of misadaptation does not imply a large excess risk: a non-optimal adaptation operator may perform almost as well as the optimal one. The paper should make this distinction clearer when discussing the practical implications of the bound.
  3. The empirical results provide limited evidence for unactionability itself. The USP DS case study shows substantial variation in the utility of fixed detector-triggered adaptation. However, this variation could also arise from detection delay, parameter settings, or the cost of resetting. Comparisons between different architectures, such as AMF and ARF, cannot separate these effects. A more direct evaluation would keep the learner, detector, and candidate adaptation operators fixed, and vary the information available to the adaptation policy. The synthetic experiments also use a single post-drift horizon and default learning parameters, leaving unclear how robust the preferred adaptation operation is to these choices.
  4. The novelty of the general theoretical results is limited. Theorem 1 assumes that no adaptation operator is optimal for every distribution in the post-drift fibre, and concludes that a policy observing only that signal must incur positive excess risk somewhere in the fibre. This is a useful formalization, but the conclusion follows quite directly from the assumption. Theorem 5 similarly applies the data-processing and Fano inequalities to the stated observation model. The main theoretical contribution therefore appears to rest on the drift-specific constructions in Theorems 2–4, whose proofs currently leave the central risk comparisons unresolved.
  5. Several inconsistencies make the constructions and experiments difficult to follow. Section 3.4 samples post-drift covariates from ([0.5,1]^2), whereas Appendix D.1 changes only (X_1) to (U[0,0.5]). For latent task inversion, Appendix B.2 changes (Z_2), while D.1 changes (Z_1). These settings are not equivalent. The notation (z) is also used for the binary detector signal, the label-noise probability, and a distributional discrepancy. In Appendix A.2, the posterior under symmetric label noise should be (z+(1-2z)\mathbf{1}(X_1+X_2>1)); the additive (z) is missing. These points need correction so that the theoretical constructions can be matched to the reported experiments.
Questions for authors
  1. Under what assumptions on the online learning rule, reset initialization, and horizon (H) do Theorems 2–4 establish that reset has strictly lower finite-horizon risk when both learners continue updating? Can the argument in Appendix A.1 also be corrected for randomized policies?
  2. How far do the information bottleneck results extend beyond policies conditioned on a single detector signal? In particular, how would access to learner state and previous detector signals affect the bound, and when does a non-zero probability of misadaptation imply meaningful excess risk?
  3. Could the authors clarify which synthetic settings generated the figures, given the differences between Section 3.4 and Appendix D.1, and between Appendices B.2 and D.1? How sensitive are the reported advantages of reset or no action to the post-drift horizon and learning parameters?
3 comments
AuthorsOpenReview ↗

Answers for rebuttal

Thank you for your detailed and constructive comments. We agree with each soundness point and have revised accordingly.

W1/Q1. Finite-horizon risk with both learners updating: The original sketches implicitly assumed the retained learner does not adapt. We now give full proofs for a specified learner: a partition count learner, which is the leaf model of a majority-class Hoeffding tree with fixed structure (Def. 3). Reset sets the counts to zero, and both learners update after adaptation. We introduce two new properties: stability and inertia (Lemmas 1–2) to explain these results. Theorem 3 (with Theorems 6–8 in App. A) states the horizons explicitly: covariate drift is unactionable for every H≥1H \geq 1, and posterior and representation drift for H above an explicit threshold set by the inertia n0n_0. Remark 8 notes that the advantage of reset vanishes as H→∞H \to \infty, so the optimal operator also depends on H and the learner’s state (this is also shown in the sensitivity analyses we added (see the next point)). The randomised-policy error in A.1 is fixed (see our reply to mwjw).

W2/Q2. Scope of the information bound: Theorem 4 holds for any finite signal space, so it covers warning/drift detectors (H(Z) ≤ log 3). We now state its scope explicitly (Remark 3): it covers policies conditioned on the detector output, which is how detector-triggered adaptation is implemented in the methods we review. Additional state information and multi-detector architectures are left to future work. On misadaptation vs. excess risk, we agree: E[δH]≥γPeE[\delta_H] \geq \gamma P_e, which says nothing when γ≈0γ \approx 0 (Remark 5). This is why our experiments report each adaptation operation’s excess risk directly.

W3/Q3. More direct empirical evidence: As you suggested, the learner, detector and candidate operators are now fixed. In the synthetic experiments, each replication forks the learner at the first detection and evaluates retain and reset on the same samples, giving paired risks (Sect. 3.4, Table 3). App. B.3 varies the horizon, ADWIN’s threshold, the learning rate and the number of pre-drift samples and show under which conditions the optimal adaptation are still preferred. Sect. 5 applies the same comparison to the USP DS streams for two learners and two detectors, against a hindsight oracle. The evidence for unactionability is that the oracle mixes retain and reset across events that all produce the same signal (action entropy 0.61–0.73), not that it outperforms fixed policies.

W4. Novelty of Theorem 1: We agree it is primarily a formalisation. The revision strengthens it to a characterisation (an if-and-only-if: a zero-excess-risk policy exists exactly when the fibre has a common optimal operator) and a quantitative bound, and the main technical contribution is now the drift-specific results with explicit horizons (Theorem 3, Lemmas 1–2).

W5. We have corrected the inconsistencies you identified. For example, Section 3.4 now uses the non-degenerate covariate shift X1∼U[0,0.5]X_1 \sim U[0, 0.5] matching App. D.1. The representation constructions in the theorem and the experiment match. z now denotes only the detector signal, the noise rate is η\eta and the posterior is now η+(1−2η)⋅1[⋅]\eta + (1 − 2\eta)·\mathbf{1}[\cdot].

Please let us know if you have any additional questions.

Thank you for the detailed response and revisions. The clarification of the information bound’s scope and the corrections to the synthetic settings help address several of my original concerns. Specifying the learner and horizon conditions, together with the paired retain/reset experiments, also represents meaningful progress. Regarding Weakness 1, resolution of the original soundness concerns depends on the revised proofs establishing the finite-horizon risk comparisons with both learners updating, together with the corrected randomized-policy bound. Regarding Weakness 3, the oracle’s action entropy is informative, but should be interpreted alongside the paired excess-risk results and their uncertainty to distinguish genuine differences in preferred operations from near-ties or sampling variability. Regarding Weakness 4, the strengthened characterization clarifies the formalization, but the substantive theoretical contribution still rests on the validity and informativeness of the drift-specific constructions under the stated assumptions.

AuthorsOpenReview ↗

Response to Comment by Reviewer ZTiY

Thank you for engaging with the revision.

W1/W4: Validity of the drift-specific constructions:

To validate the finite horizon comparisons, we introduced two lemmas (Lemmas 1–2, App. A.2):

Lemma 1 (stability). We couple the two learners on the same post-drift samples. In every cell, the retained learner’s class-count margin is always at least the reset learner’s margin plus n0n_0. When the Bayes classifier is unchanged, expected error is non-increasing in this margin, so retaining is never worse at any step. It is strictly better on every first visit to a cell, which gives RH(Ωreset)−RH(Ω0)≥(12−ηˉ) E[VH]/H>0R_H(\Omega_{\text{reset}}) - R_H(\Omega_0) \ge (\tfrac12-\bar\eta)\,\mathbb{E}[V_H]/H > 0 for all HH.

Lemma 2 (inertia). Under label flips, the retained learner makes mistakes on the first mj≥n0m_j \ge n_0 visits to each flipped cell, whereas the reset learner only makes mistakes on first visits (expected loss = 0.5). Eq. 6 gives the exact difference in risk. The horizon condition follows from a Chernoff bound on the number of visits.

Theorems 6–8 apply these two lemmas to the three drift settings, and Theorem 3 summarises them. The randomised-policy bound (App. A.1) follows by averaging: for any mixture pp, max⁡k∑mpmδmk≥1K∑k∑mpmδmk≥ϵH/K\max_k \sum_m p_m\delta_{mk} \ge \frac1K\sum_k\sum_m p_m\delta_{mk} \ge \epsilon_H/K.

W3: Action entropy vs. near-ties on USP DS:

In the real-world streams, exploring all 2k2^k action sequences is infeasible, so the oracle is obtained with a beam search (width 20). All learners and detectors are deterministic and streams order is preserved, thus there is no run-to-run randomness. We account for the remaining finite-sample variation in κ\kappa with our statistical test ZZ (App. D.2). Table 2 excludes settings where the oracle and both fixed policies do not differ.

For 40/89 settings (15 of the 24 streams, across all four detector–model pairs), the oracle statistically beats both always-retain and always-reset. This is only possible if some detections require retaining while others resetting, even though every detection produces the same signal. This is the condition of Theorem 1 (the unactionability theorem), observed on real streams. In the remaining settings, one fixed action is already near-optimal, which is consistent with our claim that unactionability can occur, not that it always does.

We agree that individual decisions may still be near-ties. We therefore treat action entropy as descriptive, not necessarily as a count of decisions that matter.

W4: Contribution:

We agree that the contribution rests on the drift-specific constructions. We acknowledge that limitation which is why we include formal proofs and empirical evidence of unactionability on three categories of drift (loss / supervised, unsupervised / covariate, and latent representation) which we believe covers the vast majority of drift settings as noted in [1]. Covariate drift is unactionable for every horizon. Posterior and representation drift are unactionable for horizons above an explicit threshold set by the learner’s inertia n0n_0.

[1] Vitor Cerqueira, Heitor Murilo Gomes, Marco Heyden, Bernhard Pfahringer, and Albert Bifet. A framework for evaluating and benchmarking concept drift detection methods. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pp. 8671–8682, 2026.

Reviewer_rg1DOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The central argument of this paper is that change detectors, in the form of a binary signal indicating concept drift, cannot uniquely determine the optimal adaptation operation. The paper presents theorems across different scenarios to demonstrate that the "unactionability" under (i) posterior drift, (ii) covariate drift, and (iii) representation drift. The theorems are proved by identifying counter scenarios where two types of distribution shifts have the same observed behavior by the change detector, such that the detector signal is not able to identify different causes of distribution drifts. In particular, Theorem 2 makes label noise indistinguishable from concept drift, Theorem 3 introduces unobservable labeling relationship Y|X change but the change detector only has access to X, and Theorem 4 applies translations to variables without information on which variable is informative vs. noise. The paper also uses information theory and data processing inequality to show the limitations of change detector. The experiments across different families of adaptation policies demonstrate substantial heterogeneity in the utility of fixed detector triggered adaptation algorithms. The paper advocates using richer signals beyond binary change detector.

Critical strengths
  1. The paper is well motivated and demonstrates the limitations of change detection in deciding adaptation algorithms.
  2. The paper provides theoretical analysis on the unactionability, and I find the definitions and theorems clear and sufficiently demonstrates the proposed argument, i.e., unactionability of change detection. I also liked the coverage of different types of concept drifts: posterior, covariate and representation.
  3. I think the key takeways of leveraging leveraging richer information to determine adaptation algorithms make sense.
Critical weaknesses

My main concern is on the unstated assumptions of the "concept drift detector" in that all the theorems require the detector to be uninformative. They introduce label noise indistinguishable from concept drift, unobservable labeling relationship Y|X change, and translations to variables without information on which variable is informative vs. noise. My main takeaway from the theorems is that if the drift detector is unreliable, the optimal adaptation policy can not be determined; therefore, I think the paper is misleading in not stating this key assumption. If the drift detector is unreliable, we do not need to go through the theorems to understand it is not informative in determining the optimal adaptation policies.

Questions for authors
  1. Please address my comment in the weakness section
  2. In the experiments, what is the accuracy of drift detection, and is there positive correlation between drift detection accuracy and adaptation performance?
4 comments
AuthorsOpenReview ↗

Answers for rebuttal

Thank you for your comments. We believe we weren't explicit enough in the paper and this caused a misunderstanding.

The detectors in our theorems are perfect, not unreliable: T is defined on population distributions, so it is an error-free detector, and in every construction the drift is real and z = 1 is a true positive in both scenarios (Remark 7). The two scenarios cannot be told apart because the detector’s input is identically distributed across them, not because the detector errs. Our main argument is that a perfectly accurate detector answers “did my input change?”, while the adaptation decision asks “which adaptation operation is optimal?”. These are different questions and the basis for unactionability.

Detection accuracy vs. adaptation performance: In the synthetic experiments, for covariate and representation drift ADWIN detects in 100% of replications and detections occur at the same instance in both scenarios of every replication, because the two scenarios share all random draws. However, the optimal operator is opposite: retain (no-action) is optimal in one scenario and reset in the other. Detection accuracy is therefore identical while adaptation utility is reversed. For posterior drift, ADWIN monitors the learner’s error, so detection rates and delays differ between scenarios, and the same reversal still holds. We have included Table 3 which now summarises the detection rate, detection delay, and the empirical risk calculation for the validations of our drift specific theorems. We also updated the Unactionability Figures (1, 4, 5) to include ticks showing the range of detection delays across the experiments.

On real streams the drift points are unknown, so detection accuracy cannot be measured. We have included an additional experiment in Section 5 where we show that within a single stream, learner and detector, the optimal action varies across detection events. We did this for the entire USP DS repository evaluating two models (Hoeffding Tree and Naive Bayes) and two detectors (ADWIN and D3). Fig. 2 shows an example on Insects-Gradual, where the oracle retains at some detections and resets at others.

Please let us know if you have any additional questions.

Our main argument is that a perfectly accurate detector answers “did my input change?”, while the adaptation decision asks “which adaptation operation is optimal?”

Even though the detector is error-free, it was not able to provide informative signal because "two scenarios cannot be told apart because the detector’s input is identically distributed across them." I wonder if the detector should be defined as uncovering the drift signal rather than any input change as any input change contain confounders as you demonstrated? Also, the title of the paper is "The Unactionability of Concept Drift Detection," not any input change detection.

AuthorsOpenReview ↗

Response to Reviewer rg1D

Thank you for your response.

The Unactionability Theorem (Thm. 1) holds for any concept drifts (changes in P(X, Y) over time) which produce an identical detection signal, but require different optimal adaptation operations. Assuming even the most informative possible concept drift detector: T(Pt,Pt+1)=1[Pt≠Pt+1]T(P_t, P_{t+1}) = \mathbf{1}[P_t \neq P_{t+1}]. which observes the full joint distribution, with no confounders and no errors. In every pair in our constructions, both scenarios are genuine concept drifts, so this ideal detector outputs z=1z=1 for both, yet the optimal adaptation differs. Theorem 1 is stated for any detector T:P×P→ZT:\mathcal{P}\times\mathcal{P}\to\mathcal{Z}, so it applies to this detector as well.

In this work we used the generally accepted definitions (Eq. 2 and Definition 1 in our paper) that Concept Drift is any change in the joint distribution over time: Pt(X,Y)≠Pt+1(X,Y)P_t(X, Y) \neq P_{t+1}(X, Y) and a change/drift detector is then simply a method which detects said changes.

Thank you for the clarification and I've revised the rating to weak accept.

Reviewer_mwjwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In online learning, concept drift corresponds to situations where the data distribution changes over time. An important aspect of online learning is then to detect concept drift and adapt the predictions accordingly. This paper addresses the question of knowing whether detection of the drift provides enough information to select appropriate model adaptation strategies. The answer to this question, carried mostly by the central theorem in the paper, is that a detection of concept drift does not necessarily indicate which adaptation operation should be performed. This idea is formalized with the notion of unactionability: different post-drift distributions can produce the same detector signal, but still require different optimal adaptation operations. The problem of unactionability is studied from different angles: a central theorem establishing unactionability, constructions involving posterior, covariate and representation drift, and finally some information-theoretic analysis. These theoretical results are complemented by synthetic experiments. The experiments reveal substantial variability in the benefits of detector-triggered adaptation, showing that more informative and potentially learnable adaptation policies could be preferable to fixed heuristics.

Critical strengths

I find the research question relevant and particularly interesting. It is often implicitly assumed that the operations of detection and adaptation to concept drift are the same, or, on contrary, that adaptation can be unique and not relying on the results of the detector. The distinction addressed by this paper, between identification of drift and appropriate adaptation, becomes there a novel question (as far as I know). In particular, I appreciate the idea of explicitly examining the information available for adaptation decisions. From that perspective, the theoretical framing is very adapted and reasonable. I appreciate the formalization of unactionability through sets of post-drift distributions associated with identical detector. The consideration of finite-horizon adaptation risk is also particularly appropriate, as different adaptation operations may involve different recovery costs. A last point I appreciate in the paper is how the theoretical discussions are complemented by an empirical analysis, here with controlled synthetic experiments. It is important, for such a specific context, that the experiments are indeed completely controlled and have no variability due to the nature of data for instance. In particular, it is noticeable that the results are not fully positive: it is indeed observed that detector-triggered adaptation can be beneficial for some architectures and detrimental for others. This "experimental no-free-lunch" is important for a nuanced discussion of the results, that is the basis of scientific integrity.

Critical weaknesses

I could not read the proofs in details (so my comments might not be fully correct) but I have two small concerns that I would like to address here: I invite the authors to comment on these and correct me.

  • Proof of Theorem 1: To me the inequality in line 695 in unclear, or more precisely, I am not sure I understand whether the provided justification is enough. As far as I understand, what is established is that each individual action has positive excess risk in at least one environment. Not that we have a same lower bound for randomzed mixtures of actions.
  • Proof of Th. 4: It is argued that a modification of the distribution of a task-relevant latent coordinate causes systematic misclassification. However, changing the marginal distribution of this coordinate does not by itself invalidate the original decision boundary. An explicit change in the conditional label distribution is needed.

Another weakness I see is the limited practical insight offered by the information-theoretic analysis. The proposed bound depends on the entropy of the optimal adaptation decision, which is not known in practical applications and may be difficult to estimate. I fail to understand whether this is meant to be a merely theoretical argument or whether the authors see any practical implication here. Moreover, the whole analysis relies on the key assumption that entropy will grow with the logarithm of the adaptation space. This assumption guarantees the asymptotic lower bound. However, I do not see whether this assumption is realistic in any practical case, in particular for the learning architectures considered in the paper. It would be important to elaborate more on this assumption, in order to establish whether it is restrictive in practice.

Questions for authors

Could you comment on my 3 main negative points:

  • Unclear statement in proof of Th 1
  • Unclear statement in proof of Th 4
  • How restrictive is the assumption made in the information-theoretic analysis?
1 comment
AuthorsOpenReview ↗

Answers for Rebuttal

Thank you for your thoughtful comments.

Q1. Theorem 1 with mixed policies: In our original formulation we did not correctly establish a lower bound for randomised / mixed policies. We have revised the Theorem (Thm. 1) and proof (App. A.1) to show that every policy has worst-case excess risk at least vHv_H, the value of a zero-sum game between the policy and the fibre. We now also show vH≥ϵH/K>0v_H \geq \epsilon_H/K > 0. The bound now holds uniformly over policies. We also added the converse: a signal admits a policy with zero excess risk if and only if its fibre has a common optimal operator.

Q2. Representation drift: You were correct that shifting the marginal of Z1Z_1 leaves the head optimal. This was a mismatch between our code implementation and what we wrote. We have revised the theorem (Theorem 8) where P(2)P^{(2)} changes the labelling rule explicitly to Y=1[Z1<δ]Y = \mathbf{1}[Z₁ < δ]. The detector signals are now exactly equal for any swap-invariant detector (e.g. ADWIN on ∣∣Z∣∣||Z||, MMD, Wasserstein). We double checked and the experiment in App. B.2 matches the Theorem.

Q3. The entropy assumption: It is a sufficient condition, not a general claim. We added Remark 4 now to explain when it holds. For example, if the operator chooses which of K parts of a model to reset and each part needs resetting independently with probability p, then c=Hb(p)c = H_b(p). The experiments in App. C simulated this case. The condition no longer holds when the optimal decision concentrates on a few operators. We now make it explicit that the fibre argument of Theorem 1 still applies. In practice, the bound tells us how many bits of signal a reliable adaptation decision needs, compared with the at most log 3 bits a binary or warning/drift detector provides.

Please let us know if you have any additional questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Additions/Revisions during Rebuttal

We first thank the reviewers for their thoughtful comments. Your comments identified legitimate errors and limitations in our proofs, and we have corrected them accordingly. We have also added a case study on the USP DS repository showing that, for a fixed learner and detector, the best action (retain or reset) varies across detection events on most streams..

Nevertheless, the claims of our paper are unchanged: a detector signal can be unactionable, the information a detector exposes limits adaptation, and fixed detector-triggered policies are suboptimal.

Below we include a summary of all the changes we made to the paper, which sections these changes fall into, and which reviewers comments they address. We have also marked the updated version of the paper with red text to indicate which content is new or revised.

Summary of Changes:

ChangeWhereRequested by
Theorem 1 holds for randomised policies with a uniform bound (vH≥ϵH/Kv_H \geq \epsilon_H/K), plus an if-and-only-if characterisation3.2, App. A.1mwjw, ZTiY
Drift-specific theorems now state the learner, the reset initialisation and the horizon H, with both learners updating after adaptation3.3 (Def. 3, Lemmas 1–2, Thm. 3), App. A.2–A.5ZTiY
Representation-drift construction corrected so that P(Y|Z) changes explicitlyThm. 8, App. A.5mwjw
Detectors in all constructions are perfect (true positives). Detection rate and delay reportedRemark 7, Table 3rg1D
Synthetic experiments use a fixed learner and detector with paired forks at detection. Constructions and experiments now match3.4, App. B, D.1ZTiY
Sensitivity to horizon, detector threshold, learning rate and pre-drift samplesApp. B.3ZTiY
Information bound: scope stated, entropy condition discussed, link to excess risk4 (Remarks 3–5)mwjw, ZTiY
Case study with fixed learner and detector, retain vs. reset vs. a hindsight oracle5, Tables 5–6ZTiY
Notation (z, η\eta, δ\delta) and inconsistencies between the main text and the appendix correctedthroughoutZTiY

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.