acceptodds
Under review as a conference paper at ICLR 2027

Strict Online Learning for Wild Streaming Data

Abstract

Real-world streaming data commonly exhibit characteristics such as concept drift, missing values, and heterogeneous features, and such data are referred to as . Existing approaches often rely on buffers that retain past instances to support model adaptation, while the implications of data retention, including privacy and data governance concerns, have received limited attention. We investigate a new problem, , in which each instance is used once for prediction and update and then discarded. This no-retention requirement makes learning challenging, as the model adapts to evolving patterns one instance at a time without destabilizing previously acquired knowledge. To address this challenge, we propose , a fast-slow online neural framework for regulating adaptation under uncertain single-instance evidence. Its slow pathway contextualizes the current observation using accumulated stream context, while its fast pathway captures instantaneous evidence. Bounded Adaptive Modulation coordinates their interaction by controlling how instantaneous evidence influences the slow representation. To stabilize single-instance learning, we introduce Online Prior Adjustment based on a prequential class prior estimated solely from previously revealed labels. Extensive experiments across 81 streaming conditions derived from 27 public datasets demonstrate that OnlineNet achieves the best aggregate predictive performance across all four cumulative metrics. It also maintains high computational efficiency while using only 0.5% of the parameters required by a competitive baseline. The source code and datasets are submitted as Supplementary Materials for Reproducibility.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
online learning, streaming data, streaming feature

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4UUaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies prequential binary classification with drift, heterogeneous features, and missing observations under a use-and-discard protocol. OnlineNet combines a slow pathway using mask-aware cumulative normalization with a fast pathway receiving current values and their availability mask. A bounded, non-attenuating multiplicative gate lets the fast representation modulate the slow representation, while training-time logit adjustment uses class counts from previously revealed labels. Experiments on over 20 datasets show that the proposed OnlineNet outperforms several baseline methods for online learning.

Critical strengths
  1. The combination of incomplete observations, evolving distributions, and prohibited instance replay is practically relevant. The prequential protocol and distinction between data retention and formal privacy are clearly stated.

  2. The implementation is compact and the method is clearly presented. Eqs (4)-(5) give an explicit gate range, and Alg. 1 updates feature statistics and class counts only after prediction and the supervised update.

Critical weaknesses
  1. This work is largely an incremental design based on existing methods, and the novelty relative to existing no-buffer online learners needs sharper positioning. Section 3's test-then-train, single-instance update protocol is already compatible with most evaluated baselines, and Table 5 explicitly marks HBP, MODL, HI2, and several classical methods as not retaining prior instances. Therefore, the contribution is principally a particular architecture for this setting rather than the no-retention setting itself. Welford normalization and Menon et al.'s logit adjustment are established ingredients, leaving the gate as the main technical addition. Please explain its distinction from standard conditional feature modulation and residual gating, and isolate the benefit over a compact masked-input MLP with the same causal normalization and prior-adjusted loss. This would make the architectural contribution more convincing.

  2. The experiments are not sufficient, because this paper targets a very challenging and practical setting of learning from wild and non-stationary data. However, the experiments are conducted on synthetic datasets with self-crafted feature missing. Section 5.1 removes features independently with a fixed probability and ensures at least one feature remains observed; all tasks are binary, with at most 50,000 instances and majority fractions below 77%. This does not test correlated sensor failures, missingness linked to values or labels, changing availability mechanisms, feature birth/death, severe evolving class imbalance, or co-evolution of features and labels. Experiments on more large-scale real-world datasets with millions of instances, self-evolving distributions, and severe class imbalance should be conducted to evaluate the effectiveness of the proposed architecture. For example, in financial fraud detection (a binary classification task), the fraudulent patterns are continually evolving, the fraudulent rate keeps changing over time, feature missing is common, and the class imbalance problem is a major concern. Such a realistic setting should be considered in this paper, since this paper studies streaming learning with concept drift, missing values, and heterogeneous features. Moreover, the classification task considered is binary, and there are many scenarios with multiple classes. It is better if the authors can provide these experimental results.

  3. The contribution of each component in the OnlineNet should be further evaluated. As shown in Tab.13, in many cases, removing certain components can lead to performance improvement. Moreover, the contribution of the proposed modules is not significant, and the current ablations do not fully isolate the proposed fast-slow mechanism from normalization and model capacity. Tab. 3 (average results) shows the largest balanced-accuracy drop when removing causal normalization, from 77.422 to 73.427, whereas removing the fast pathway gives 76.772. The first result establishes that scaling matters, but it does not establish that two complementary pathways are necessary.

  4. Cumulative context and prior estimates can become stale under the very drift the method targets. Appendix C.1 uses all-history Welford means and variances, and Eq. (7) estimates class frequencies from all previously revealed labels. After a long stable period, a new regime has diminishing influence on these statistics; a formerly rare class can remain strongly adjusted even when its current frequency changes. The cumulative scores and sliding-window plots in this paper do not specifically quantify this failure mode.

Questions for authors

see weaknesses

6 comments

W1.1: Positioning and Contributions

Thank you for your comment. We provide the following clarifications.

(a) Relationship to Existing No-Buffer Online Learners. As discussed in Appendix A of the original paper, privacy protection and data governance motivate the formulation of constraints governing the use and retention of streaming data. However, existing paradigms lack a unified specification of these requirements. Standard online learning emphasizes sequential prediction and incremental model updates but does not prescribe whether historical instances should be retained. Buffer-free learning excludes replay buffers but does not constrain historical-sample access or batch size. For example, in online time-series forecasting, historical samples may be repeatedly accessed as temporal contexts, whereas online continual learning may employ mini-batch processing and multiple training epochs. Therefore, we characterize Strict Online Learning through explicit constraints on instance usage, update frequency, and historical data retention.

(b) Test-Then-Train, No-Retention, and Single-Instance Updates. We clarify that the prequential test-then-train procedure is a standard evaluation protocol in online learning. Single-instance updates and the non-retention of prior instances adopted by some of the evaluated baselines are algorithmic choices for incremental adaptation. In particular, discarding prior instances avoids repeated reliance on potentially outdated observations as concepts evolve. We emphasize that we have never claimed that these practices themselves constitute a novel learning setting or a methodological contribution.

(c) Contributions. Our contribution extends beyond a particular architectural design. We systematically investigate the learning challenges posed by wild streaming data under the strict online learning constraints described above. As discussed in the Introduction, this setting introduces a distinct stability–plasticity dilemma under uncertain single-instance evidence. Each update relies on potentially incomplete observations, while predictive relationships evolve over time. Without access to historical instances for replay, potentially disruptive updates cannot be reassessed or corrected by revisiting those observations. Moreover, evolving class frequencies can introduce bias. These coupled challenges constitute the central learning problem investigated in this work.

W1.2: Components and Gate

Thank you for your comment. We provide the following clarifications and additional experiments. We will incorporate the corresponding changes into the revised paper and highlight them in blue.

Components

We do not claim Welford normalization or logit adjustment to be novel contributions. We adapt these components to our setting. For Welford normalization, we develop a mask-aware causal normalization scheme based on Welford's algorithm to accommodate feature missingness. For logit adjustment, since the global class prior is unavailable in streaming data, we estimate class priors prequentially from previously revealed labels and use these estimates to adjust logits in the training objective.

We emphasize that the primary technical contribution of this work is OnlineNet, a lightweight neural Fast–Slow framework. Although relatively simple, its architecture reflects a deliberate design choice for learning from wild streaming data, where computational efficiency and generalizability are important. OnlineNet has fewer than 0.5% as many parameters as a state-of-the-art model while achieving strong predictive performance across diverse datasets and domains.

BAM vs. Standard Conditional Feature Modulation and Residual Gating

(a) Difference. The key distinction lies in how BAM integrates instantaneous evidence into the slow representation. Standard conditional feature modulation typically applies feature-wise scaling and shifting, whereas residual gating controls the contribution of a separately computed residual representation. In contrast, BAM directly applies multiplicative, bounded, and non-attenuating modulation to the slow representation. Weak responses from the fast pathway keep the gate near its multiplicative identity, while stronger responses allow amplification. By preventing attenuation, BAM protects accumulated contextual information and keeps the slow representation as the basis of the fused representation, while allowing the fast pathway to regulate its feature responses.

(b) Additional Experiments. We compare BAM with FiLM-style modulation and Gated Residual Fusion, which represent standard conditional feature modulation and residual gating, respectively. All other model components and training settings remain unchanged. The averaged results are shown below.

ModelAccuracyAUROCAUPRCBalanced Accuracy
FiLM-style Modulation77.4581.5079.0875.99
Gated Residual Fusion78.2182.5878.7376.50
BAM88.0093.7491.2186.97

W1.3: OnlineNet vs. masked-input MLP variant

Thank you for your comment. We compared OnlineNet with a compact masked-input MLP variant, referred to as OnlineMLP-OPA, using the same causally normalized inputs, availability masks, and OPA prior-adjusted loss. The results are reported below. Each cell shows the OnlineMLP-OPA score, followed by the difference (OnlineMLP-OPA − OnlineNet) in percentage points.

DatasetAccuracyAUROCAUPRCBalanced Accuracy
wdbc78.38 (−4.50)86.49 (−2.80)81.07 (−3.23)74.34 (−6.37)
spam86.72 (−4.27)93.68 (−2.36)88.95 (−3.56)83.12 (−6.26)
diabetes_f64.58 (−2.43)60.43 (−3.71)43.75 (−5.62)51.76 (−4.20)
spambase77.53 (−2.99)83.79 (−3.18)77.29 (−3.89)75.37 (−3.66)
LUdata63.12 (−3.23)70.82 (−3.22)70.78 (−3.39)63.12 (−3.18)

W2.1: Datasets and Experimental Setup

Thank you for your comment. We would like to provide the following clarifications.

(a) Dataset Coverage and Experimental Breadth. We clarify that our evaluation is not limited to synthetic datasets. Our original experiments cover 27 datasets, comprising 15 real-world datasets spanning 11 application domains and 12 synthetic datasets generated using MOA. Among the real-world datasets, four are time-series datasets, namely LUdata, spam, weather, and elec. The datasets collectively cover numerical, categorical, mixed-type, and binary-valued features. Across all datasets, the number of instances ranges from 569 to 50,000, and the number of features ranges from 3 to 499. The majority-class percentages range from 50.32% to 76.19%. The synthetic datasets further incorporate controlled abrupt and gradual concept drifts with predefined transition positions. The 27 datasets evaluated at three observation probabilities yield 81 dataset–availability conditions. This enables systematic evaluation across different data scales, feature types, and feature availability levels.

(b) Feature Availability Protocol. We adopt a controlled feature-masking protocol, following prior online learning studies [1–4], to systematically evaluate OnlineNet's robustness under different degrees of feature missingness. For each instance, each feature is independently sampled as observed with probability pp. Specifically, we consider three feature observation probabilities, $p \in \{0.25, 0.50, 0.75\} $, corresponding to nominal missing rates of 75%, 50%, and 25%, respectively. If all features are sampled as unavailable, one feature is randomly retained. This protocol enables controlled and reproducible comparisons across different feature availability levels.

(c) Scope of Research. Our work focuses on strict online learning for wild streaming data, particularly the challenges of concept drift, varying feature availability, and heterogeneous features. Accordingly, our evaluation examines different levels of feature missingness and both abrupt and gradual concept drifts across real-world and synthetic datasets. The scenarios raised by the reviewer, including correlated sensor failures, missingness mechanisms dependent on feature values or labels, dynamically changing availability mechanisms, feature birth and death, severe and evolving class imbalance, and the co-evolution of features and labels, involve additional problem dimensions and assumptions that extend beyond the specific scope of wild streaming data investigated in this work.

[1] Yi He et al., Online Learning in Variable Feature Spaces with Mixed Data, ICDM, 2021.

[2] Rohit Agarwal et al., Aux-Drop: Handling Haphazard Inputs in Online Learning Using Auxiliary Dropouts, TMLR, 2023.

[3] Antonios Valkanas et al., MODL: Multilearner Online Deep Learning, AISTATS, 2025.

[4] Yulin Wang et al., Online Learning for Wild Streaming Data with Delayed Feedback, SIGKDD, 2026.

W2.2: Extreme Class Imbalance

Thank you for your comment. As detailed in our response to W2.1, we have made efforts to ensure broad experimental coverage across challenging streaming conditions. We provide the following further clarifications and additional experimental results.

(a) Scope of This Work. We agree that evaluating our approach on larger-scale financial fraud detection datasets is valuable. However, developing methods specifically targeting extreme class imbalance is beyond the scope of this work. We focus on strict online learning for wild streaming data, aiming to balance adaptation and knowledge retention under intermittent feature missingness and non-stationarity without retaining historical instances. Although OPA accounts for class-frequency effects during online learning, it is not intended as a dedicated method for extreme class imbalance. This problem is commonly studied as online class imbalance learning [1–3]. Integrating specialized imbalance-aware learning strategies into the strict no-retention setting is therefore an important direction for future research.

(b) Additional Experiments. To provide further empirical evidence, we conduct additional experiments on BAF Variant IV, a synthetic bank account fraud benchmark based on anonymized real-world data. It contains 1,000,000 instances, including 11,030 fraudulent cases (1.103%), and 30 input features. The records span eight months and are processed chronologically, with an approximately 5.89% native feature missingness rate. We evaluate two feature availability settings, pp=0.5 and pp=1.0, where the former randomly masks 50% of originally observed feature values, whereas the latter preserves the original missingness. The comparative results are reported below, with BA denoting Balanced Accuracy.

Methodpp=0.5 AUROCpp=0.5 BApp=1 AUROCpp=1 BA
OnlineNet68.3750.1278.6051.39
MODL50.0050.0050.0050.00
Aux-Drop49.5850.0048.6050.00
OLIFL50.0149.9950.0249.98
OLVF51.3850.0940.4550.04
ORV$^3$V45.8650.0045.0450.00

These results demonstrate OnlineNet's scalability to million-instance streams and its stronger ranking discrimination. They also highlight the remaining challenges in minority-class classification.

[1] Aristotelis Chrysakis et al., Online Continual Learning from Imbalanced Data, 2020, ICML.

[2] Dianlong You et al., Online Learning From Incomplete and Imbalanced Data Streams, 2023, IEEE TKDE.

[3] Jiahang Tu et al., Adaptive Learning in Imbalanced Data Streams With Unpredictable Feature Evolution, 2025, IEEE TKDE.

W2.3: Multi-class Classification

Thank you for your comment. We provide the following clarifications and additional experimental evaluations.

(a) Fairness Considerations. Some baseline methods were originally designed specifically for binary classification. Although these methods can, in principle, be extended to multi-class classification through a One-vs-Rest (OvR) strategy, doing so would require adapting their underlying online learning and prediction mechanisms. Since the original papers do not specify how these methods should be adapted to multi-class settings, we limited our original evaluation to binary classification to avoid introducing additional design choices that could compromise the fairness of comparisons.

(b) Additional Experiments. We conduct additional comparative experiments on Asfault, a five-class streaming dataset. For fair comparison, we select only neural network-based baselines from the same model family as OnlineNet. The following table reports the average performance at different feature availability rates.

ModelAccuracyBalanced AccuracyAUROCAUPRC
HBP_multiclass61.1631.5771.8532.10
HI²_multiclass81.7661.1993.3469.01
Aux-Drop_multiclass55.5823.8372.3629.13
OnlineNet_multiclass84.3375.7296.1578.76
Reviewer_vKSVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the issue that existing methods rely on retaining historical instances. It proposes a strict online learning approach specifically for wild stream data. This task requires that each instance be used only once for prediction and update. To address this challenge, the authors, based on the complementary learning system theory, proposed OnlineNet. It utilizes a slow branch that normalizes based on historical cumulative statistics, and a fast branch that directly receives the current input and missing mask, respectively, to model the historical background and real-time information. The method was evaluated on 81 streaming settings derived from 27 datasets, demonstrating excellent overall classification performance and computational efficiency.

Critical strengths
  1. The author investigated how the model can better handle feature absence, heterogeneous inputs and concept drift in streaming data without retaining historical instances. This is highly applicable to real-world streaming data.
  2. The author reported the parameter count and running time of OnlineNet. In terms of efficiency and parameter count, it has shown excellent performance, which is very significant for practical deployment.
  3. The experimental design is quite comprehensive, covering both real data and synthetic data, and incorporating different degrees of feature absence as well as concept drift situations such as mutations and gradual changes. It examines the model's performance under various streaming data conditions. At the same time, the authors evaluate the method from multiple dimensions including prediction performance, online performance changes, and computational efficiency, thereby fully demonstrating the applicability and practical value of the method.
Critical weaknesses
  1. Currently, there are numerous studies in the field of streaming data mining that focus on non-replayable samples. The author has taken this as a new issue, but the innovation is insufficient. The author should further elaborate on the differences between this and existing research.
  2. The author mentioned in the abstract that the model needs to avoid disrupting previous knowledge. However, this point is not directly verified in the article. The existing performance demonstrations mainly reflect the predictive ability and adaptation effect, and cannot yet demonstrate that the historical knowledge has been maintained during continuous updates.
  3. Apart from the overall classification indicators, could the author provide an analysis of the changes in branch representations, the relationship between gated responses and sample reliability, or an analysis of the perturbations caused by updates under mismatched conditions, in order to support the proposed stability explanation?
  4. The author verified the model's performance on multiple datasets. As mentioned in the third section, these tasks all adopted a binary classification setup. However, in real streaming data, as the number of categories increases, the relationships between categories, frequency imbalance, and prior category estimation may become more complex. The applicability of the model in multi-classification tasks has not yet been demonstrated.
Questions for authors
  1. The paper emphasizes the need to avoid disrupting previous knowledge when adapting to new data. However, currently, the main reports focus on the overall prediction performance and the sliding window error rate. The authors should separately test the adaptability to new concepts and the ability to retain previous knowledge.
  2. Currently, the sample size of the dataset is relatively small. How does the model perform on large-scale time series tasks?
6 comments

W1: Comparison with Existing Studies

Thank you for your comment. We would like to clarify that we have never claimed the replay-free setting itself constitutes a novel learning setting or methodological contribution. We provide the following clarifications, which will be incorporated into the revised paper with the changes highlighted in blue.

(a) Existing Replay-Free Studies. Existing replay-free studies may differ from our setting in two aspects, namely data processing granularity and the number of training epochs. Specifically, some methods process multiple instances together in mini-batches [1–4], and several of these methods also perform multiple training epochs over the available data [1, 3, 4].

(b) Motivation for Strict Online Learning. As discussed in Appendix A of the original paper, privacy protection and data governance motivate the formulation of explicit constraints governing the use and retention of streaming data. However, existing learning paradigms do not necessarily provide a unified specification of these requirements. In particular, standard online learning emphasizes sequential prediction and incremental model updates but does not prescribe whether historical instances should be retained. Accordingly, we characterize Strict Online Learning through explicit constraints on instance usage, update frequency, and historical data retention.

(c) Research Gap. The learning challenges arising from wild streaming data under the strict online learning constraints described above remain insufficiently explored. As discussed in the Introduction, this setting introduces a distinct stability–plasticity dilemma under uncertain single-instance evidence. Each update relies on potentially incomplete observations, while the underlying predictive relationships evolve over time. Without access to historical instances for replay, potentially disruptive updates cannot be directly reassessed or corrected by revisiting the affected instances. Moreover, evolving class frequencies can introduce additional bias through sequential single-instance updates. These coupled challenges constitute the central learning problem investigated in this work.

[1] Seohyeon Cha et al., Task-Agnostic Federated Continual Learning via Replay-Free Gradient Projection, 2026, TMLR.

[2] Qiuyang Li et al., ReCATS: Replay-Free Continual Anomaly Detection for Non-Stationary Multivariate Time Series, 2026, SIGKDD.

[3] Huancheng Chen et al., Replay-Free Continual Low-Rank Adaptation with Dynamic Memory, 2026, CVPR.

[4] Jiayu Zhang et al., PI-CCA: Prompt-Invariant CCA Certificates for Replay-Free Continual Multimodal Learning, 2026, ICLR.

W2 & Q1: Adaptation and Knowledge Retention

Thank you for your comment. We clarify the evidence from our original experiments and present additional quantitative analyses of adaptation and knowledge retention. We will incorporate the corresponding changes into the revised paper and highlight them in blue.

Clarifying Existing Evidence

(a) Adaptation. Our experiments in Section 5.2.2 and Appendix E.2 illustrate OnlineNet's real-time performance as instances continuously arrive. Overall, OnlineNet maintains relatively low error levels, exhibits smaller increases in prediction error than baseline methods when encountering new concepts, and quickly recovers to a low error level as learning progresses. These observations provide implicit evidence of its ability to adapt to evolving data streams.

(b) Knowledge Retention. Under the feature availability settings in Section 5.1, features may remain unavailable across consecutive instances and reappear later, while the model continues to update using other observed features. This setting implicitly reflects the preservation and reuse of previously acquired feature-specific knowledge. Furthermore, competitive performance on time-series datasets such as elec and weather, both of which exhibit recurring temporal patterns, provides additional evidence that previously acquired knowledge remains useful when similar patterns recur.

Additional Quantitative Analyses

We divide the data stream into five consecutive intervals with alternating feature availability: $A(D_1) \rightarrow B(D_2) \rightarrow A(D_3) \rightarrow $$ B(D_4) \rightarrow A(D_5),where, whereA$ and$B$ are two disjoint feature subsets. Specifically, only$A$ is available in$D_1$, D3D_3, and$D_5$, while only$B$ is available in$D_2$ and$D_4$. The experimental results are reported in terms of interval-wise Balanced Accuracy (BA):

IntervalOnlineNetOnlineNet-RetrainingOLVFOLIFLAux-DropOVFMHI$^2$MODLHBP
D1D_188.2488.2455.6647.3173.2885.5678.7186.6647.92
D2D_292.4391.7367.2150.4289.2090.3391.7386.8376.70
D3D_394.3887.0465.2139.5284.4487.1492.1790.4474.60
D4D_494.2183.7977.0750.0092.4692.7093.0494.2194.21
D5D_596.3586.9460.8439.1284.4492.1495.9793.4790.58

(a) Adaptation to Newly Available Features. We evaluate adaptation using$D_2$, where feature subset$B$ becomes available for the first time after the initial$A$-only interval. OnlineNet achieves a BA of 92.43% in this interval, outperforming all competing methods.

(b) Knowledge Retention: Performance Across Reappearing Feature Subsets. We evaluate knowledge retention by examining performance on the same feature subsets across their successive appearances. For subset$A$, the BA increases from 88.24% in$D_1$ to 94.38% in$D_3$, and further to 96.35% in$D_5$. Similarly, for subset$B$, the BA increases from 92.43% in$D_2$ to 94.21% in$D_4$. These results indicate that OnlineNet maintains and further improves its predictive performance as previously encountered feature subsets reappear.

(c) Knowledge Retention: Benefits of Preserving Historical Learning States. To further examine the benefits of retaining historical learning states, we compare OnlineNet with OnlineNet-Retraining, which is reinitialized at the beginning of each interval and trained independently using only the current interval's data. In$D_3$, D4D_4, and$D_5$, OnlineNet achieves BA scores of 94.38%, 94.21%, and 96.35%, respectively, compared with 87.04%, 83.79%, and 86.94% for OnlineNet-Retraining. The corresponding gains are 7.34, 10.42, and 9.41 percentage points. These results demonstrate the predictive benefits of preserving and exploiting previously acquired knowledge when the corresponding feature subsets reappear.

W3: Further Quantitative Analyses

Thank you for your comment. As detailed in our response to W2 & Q1, we have added quantitative analyses of knowledge retention (stability) and adaptation (plasticity). Here, we analyze branch representations, gated responses, and prediction perturbations to further substantiate our explanation of stability.

We reuse the A−B−A−B−AA-B-A-B-A data stream, where AA and BB are two disjoint feature subsets, with AA available in D1,D3D_1, D_3, and D5D_5 and BB available in D2D_2 and D4D_4. We reserve 20% of the original data as an independent probe set through stratified sampling and use the remaining 80% to construct the online stream. The probe set is held out from online training and used exclusively for the following analyses. For each probe sample, we mark all features in BB as missing, forming a fixed AA-only probe set. We save the complete model state after each interval as checkpoint CiC_i. Using the same probe inputs, we then extract the Fast, Slow, and Modulated Representations from the final paired layer.

(a) Changes in Branch Representations. We use Normalized Representation Drift (NRD) to quantify the magnitude of changes in individual sample representations, with lower values indicating less drift. Taking C1C_1 as the reference checkpoint, we report the NRD values of the Fast and Slow pathways at each checkpoint.

BranchC1C_1C2C_2C3C_3C4C_4C5C_5
Fast0.0000.3850.5620.6330.743
Slow0.0000.3210.3980.3780.470

The Slow Representation exhibits consistently lower NRD than the Fast Representation across subsequent checkpoints, indicating greater stability of the Slow Pathway.

(b) Gated Responses. We interpret the reviewer's reference to gated responses as referring to the modulation gate gℓ,tg_{\ell,t} in our Bounded Adaptive Modulation (BAM) mechanism. The Fast Pathway generates this gate, which provides dimension-wise coefficients to modulate the Slow Representation according to instantaneous evidence. We clarify that these coefficients do not directly quantify sample reliability.

To characterize the gate and its modulation effect, we report the mean modulation gate value (Mean Gate), which summarizes the average gate coefficient, and the relative modulation magnitude (Rel. Mod.), which measures the relative change in the Slow Representation induced by BAM. The latter is defined as M=∥h−s∥2/(∥s∥2+ϵ)M=\|h-s\|_2/(\|s\|_2+\epsilon), where ss and hh denote the Slow Representation before and after modulation, respectively. The results are as follows:

CheckpointMean GateRel. Mod.
C1C_11.2240.553
C2C_21.2480.580
C3C_31.2720.617
C4C_41.2710.622
C5C_51.2510.608

The non-zero modulation magnitudes across all checkpoints confirm that BAM actively modulates the Slow Representations. Both quantities remain consistent across checkpoints, indicating that the modulation mechanism maintains a stable response during successive updates.

(c) Prediction Perturbations under Mismatched Updates. We conduct an additional experiment to quantify prediction perturbations caused by online updates under mismatched feature-availability conditions. Specifically, during the two BB-only intervals, D2D_2 and D4D_4, the model is updated using feature subset BB, while the fixed probe set contains only features from subset AA. This setting allows us to assess how learning BB affects predictions on previously encountered AA-only inputs.

We compare predictions on the fixed AA-only probe set before and after each BB-only interval, corresponding to checkpoint transitions C1→C2C_1 \rightarrow C_2 for D2D_2 and C3→C4C_3 \rightarrow C_4 for D4D_4. We report the Prediction Flip Rate (PFR), defined as the proportion of probe samples whose predicted labels change between checkpoints, alongside the corresponding Balanced Accuracy (BA).

IntervalPFRBA (Before$\rightarrow$ After)
D2D_22.14%89.54$\rightarrow$92.66
D4D_43.57%92.12$\rightarrow$92.57

The PFR remains low in both intervals, while BA increases after each BB-only update. These results indicate that updates under mismatched conditions induce only limited prediction perturbations while preserving classification performance.

Reviewer_eX46OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces strict online learning for 'wild' streaming data, where wild refers to the heterogeneous nature, any missing features and concept drift of the data. Another important component in its scope is the restriction is that the instances in the data can not be stored once it has been used once to predict and update the model in order to comply with privacy regulations that might demand it, although no privacy guarantee is provided. This means that only model parameters and summary statistics are retained, which is how it distinguishes itself from buffer and replay-based techniques. The proposed method is OnlineNet, which is a two-pathway MLP. The slow pathway receives input normalized with statistics from past observations that are updated after each step. The fast pathway utilizes the raw zero-filled input concatenated with the availability mask. BAM passes these representations through non-attenuating multiplicative gating and thus the fast pathway can only amplify the slow pathway. The method is tested against 8 baselines across 27 datasets. OnlineNet reports the best average Accuracy, AUROC, AUPRC and Balanced Accuracy at every value of p with an average rank of about 1.0–1.2.

Critical strengths

OnlineNet is a small and simple approach relative to the heavier methods that it is compared against, using only about 104K parameters and a two-layer MLP design. The paper also consistently reports gains over the strongest baseline across different settings. The authors also release the code and the associated implementation details for reproducibility.

Critical weaknesses

The major novel contribution claimed by OnlineNet is the fast and slow pathway mechanism, however Table 3 in the ablation section seems to show much of the gain comes from causal normalization. In particular, the fast pathway seems to add little. On german, removing it shows higher AUROC than the whole of OnlineNet and this variant also seems to cost at most 0.65 points on any averaged metric. It is not evident if the baselines also received the same normalized data and if they perform better with access to it, it implies the methodological contribution is thinner than presented. While it is unclear if any of the baselines are tested beyond their default configurations, OnlineNet uses different learning rates and depths for each dataset and no cross-transfer analysis (e.g. results with a single configuration) is provided.

Questions for authors

In addition to addressing the concerns in the weakness section, these are a few clarification questions that I would like to ask the authors:

How were OnlineNet’s learning rates and depths chosen for each dataset? Did you compare your approach against a plain MLP that also receives the same causal normalized data and mask input? What actually changes across seeds for the real datasets? While the paper presents investigating strict online learning as a contribution, what additions are made beyond the settings that the baselines already follow?

6 comments

W1: Component Contributions and the Fast–Slow Architecture

We thank the reviewer for this helpful comment. We clarify the original ablation findings and report the results of three additional experiments. We will incorporate the following discussion into the revised paper, with the changes highlighted in blue.

Clarification of the Original Ablation Results

As shown in the original ablation results, causal normalization contributes more substantially to average performance, whereas the contribution of the fast pathway varies across datasets. On elec, removing the fast pathway leads to performance drops of 1.36, 1.57, 1.63, and 1.48 percentage points in Accuracy, AUROC, AUPRC, and Balanced Accuracy (BA), respectively. On german, removing the fast pathway improves AUROC and AUPRC while reducing Accuracy and BA. These differences may partly reflect variations in dataset characteristics and realized feature-availability patterns, as discussed in Appendix E.4, as well as our use of largely shared hyperparameter settings across datasets to promote generality, rather than extensively optimizing the model separately for each dataset. Nevertheless, the complete OnlineNet achieves higher average performance than its fast-pathway-ablated variant across all four metrics, indicating the overall benefit of incorporating the fast pathway.

Additional Analysis of the Fast–Slow Architecture

We conducted three additional experiments to further examine the design and effectiveness of the proposed fast–slow architecture.

(a) Adaptability. An important advantage of the proposed fast–slow architecture is its flexibility to accommodate different backbone encoders according to domain-specific data characteristics. For real-world time-series data, where temporal dependencies are important, we replace the slow-pathway MLP with a standard two-layer gated recurrent unit (GRU), keeping the fast pathway and all other components unchanged. This variant is referred to as OnlineNetGRU. The results demonstrate the effectiveness of adapting the slow-pathway backbone to the characteristics of the data and are reported below in terms of Balanced Accuracy (BA), with each cell showing the OnlineNetGRU score followed by the difference (OnlineNetGRU − OnlineNet).

Observation Probabilityelecweather
25%69.96 (+1.67)59.77 (+1.70)
50%75.71 (+2.84)64.94 (+1.03)
75%79.60 (+2.45)69.35 (+0.62)

(b) Asymmetric Roles. To examine whether the two pathways have distinct functional roles, we construct OnlineNetGRUSwapped by swapping the GRU and MLP encoders in OnlineNetGRU, placing the GRU in the fast pathway and the MLP in the slow pathway. The two variants have identical numbers of trainable parameters and use the same training settings. The results on elec are reported below, with each cell showing the OnlineNetGRUSwapped score followed by the difference (OnlineNetGRUSwapped − OnlineNetGRU).

Observation ProbabilityAccuracyAUROCAUPRCBA
25%67.05 (−4.09)73.69 (−3.54)68.63 (−3.15)65.39 (−4.57)
50%71.66 (−4.75)79.78 (−4.25)75.71 (−4.33)70.57 (−5.14)
75%76.18 (−3.92)84.49 (−3.53)80.99 (−4.14)75.47 (−4.13)

These results support the rationale for the role-specific design of the proposed fast–slow architecture: the two pathways serve complementary but non-interchangeable roles, and their effectiveness depends on aligning encoder characteristics with their distinct functional roles. Specifically, the slow pathway is designed to construct history-conditioned representations, whereas the fast pathway focuses on extracting information from current observations and their availability patterns.

(c) Fast-Pathway Ablation in OnlineNetGRU. Building on the GRU-based configuration, we further examine whether the fast pathway continues to provide complementary information when the slow pathway employs a different backbone. Specifically, we construct OnlineNetGRUNoFast by removing the fast pathway from OnlineNetGRU. The results on elec are reported below, with each cell showing the OnlineNetGRUNoFast score followed by the difference (OnlineNetGRUNoFast − OnlineNetGRU).

Observation ProbabilityAccuracyAUROCAUPRCBA
25%70.82 (−0.32)76.81 (−0.42)71.25 (−0.53)69.65 (−0.31)
50%75.26 (−1.15)82.59 (−1.44)78.27 (−1.77)74.37 (−1.34)
75%79.63 (−0.47)87.44 (−0.58)84.04 (−1.09)79.09 (−0.51)

These results provide further evidence that the fast pathway contributes complementary instantaneous information beyond the temporal representations learned by the GRU-based slow pathway, demonstrating that its contribution is not restricted to the original MLP-based architecture.

If you are also interested in how the proposed fast–slow architecture supports knowledge retention and adaptation, please refer to our responses to Reviewer 2's comments W2 & Q1 and W3.

W2: Normalization

We thank the reviewer for this helpful comment. We clarify that online normalization is not an inherent requirement of online learning, but an algorithm-specific design choice. Different methods adopt input-processing strategies suited to their respective architectures and learning mechanisms. We conducted the following two sets of additional experiments to provide empirical evidence that indiscriminate application of online normalization can impair model performance. We will incorporate the following discussion into the revised paper, with the changes highlighted in blue.

(a) Analysis of Baselines with Causal Normalization. We applied the same Causal Normalization used in OnlineNet to Aux-Drop, HBP, and MODL, while keeping their original learning mechanisms unchanged. We excluded HI² from this comparison because it already employs its own online normalization mechanism. The following tables report the results on wbc. Here, pp denotes the feature observation probability. Each cell shows the normalized score followed by the difference (normalized − original) in percentage points.

pp=0.75

ModelAccuracyAUROCAUPRCBalanced Accuracy
Aux-Drop90.84 (+5.46)96.09 (+5.63)93.42 (+5.65)87.90 (+4.11)
HBP81.69 (+0.06)87.00 (+5.62)80.44 (+0.08)80.32 (−0.20)
MODL94.42 (+2.26)98.45 (+1.30)97.07 (+1.82)93.48 (+2.35)

pp=0.25

ModelAccuracyAUROCAUPRCBalanced Accuracy
Aux-Drop76.39 (−5.70)83.92 (−1.10)76.86 (−3.71)66.55 (−10.52)
HBP61.09 (−14.19)55.87 (−15.79)38.50 (−28.79)50.94 (−17.55)
MODL87.41 (−0.34)93.22 (+2.12)89.74 (+1.51)83.41 (−1.09)

The results indicate that the effectiveness of Causal Normalization varies with the observation probability. At pp=0.75, it generally benefits the tested baselines, whereas at pp=0.25, it degrades the performance of most baselines. This may be because historical statistics provide useful context when observations are abundant, whereas relying solely on normalized representations under sparse observations may compromise the preservation of valuable instantaneous information.

(b) Analysis of Path-Specific Causal Normalization. In OnlineNet, we specifically apply Causal Normalization to the slow pathway to construct history-conditioned representations, while allowing the fast pathway to directly process raw observations to preserve instantaneous feature-level information. This asymmetric design aims to capture complementary information from historical statistics and current observations. To further validate this design, we applied Causal Normalization to the fast pathway while keeping the slow pathway unchanged. The following table reports the results of this variant on diabetes_f, with each cell showing its score followed by the difference (variant − OnlineNet) in percentage points.

Observation ProbabilityAccuracyAUROCAUPRCBalanced Accuracy
25%64.71 (−2.30)61.54 (−2.60)43.16 (−6.21)52.21 (−3.75)
50%65.62 (−3.23)67.69 (−2.38)50.06 (−4.93)56.89 (−4.09)
75%70.31 (−0.76)73.55 (−1.38)59.10 (−1.68)63.52 (−1.66)

The observed degradation further supports the path-specific use of Causal Normalization in OnlineNet, as applying it to both pathways may compromise the fast pathway's ability to preserve instantaneous feature-level information and weaken the complementarity between the two pathways.

W3: Baseline Configurations and Cross-Dataset Generalization

Thank you for your comment. We provide the following clarifications and additional experimental results, which will be incorporated into the revised paper with the changes highlighted in blue.

Baseline Configurations

We clarify that all baseline methods were configured following their original papers and publicly available implementations. The table below lists representative candidate values considered in the hyperparameter searches.

BaselineHyperparameter Search Space
OLIFLThreshold selection option$\in\{0,1\};trade−offparameter(foroption1); trade-off parameter (for option 1)\in\{5\times10^{-6},10^{-5},0.01,0.02,0.05,0.1,0.2,0.5,1,2\}$
OLWDFDecay rule$\in \{2,3,4\};errorcontributionrate; error contribution rate\in \{0.001,0.005,0.01,0.02,0.05\};batch−sizedenominator; batch-size denominator\in \{8,10,20\}$
ORF³VForest size$\in \{3,5,10\};replacementinterval; replacement interval\in \{5,10\};replacementstrategy; replacement strategy\in \{\mathrm{oldest},\mathrm{random}\};updateweight; update weight \in \{0.01,0.1,0.3,0.9\}$
Aux-DropLearning rate$\in \{0.001,0.005,0.01,0.1\};dropoutprobability; dropout probability\in \{0.3,0.5\};auxiliary−layerwidth; auxiliary-layer width\in \{100,200,300,500,3000\}$
⋯\cdots⋯\cdots

We note that the supplementary material accompanying our original submission already provides the implementation details, configuration files, and selected hyperparameter settings for individual datasets.

Evaluation with Unified Hyperparameters

Although we made dataset-specific adjustments for four datasets, the remaining 23 out of 27 datasets (85.2%) used an identical hyperparameter configuration. To further assess the generality of this configuration, we reran the remaining four datasets (wbc, spambase, a8a, and spam) using a unified learning rate of 10−310^{-3} and two Fast–Slow layer pairs, while keeping all other hyperparameters unchanged. Each cell reports the unified-configuration score, followed by its difference from the original score (unified − original).

DatasetObservation ProbabilityAccuracyAUROCAUPRCBalanced Accuracy
wbc25%87.90 (−0.80)93.17 (−0.74)89.20 (−0.62)85.38 (−1.08)
50%93.71 (−0.42)97.64 (−0.17)95.71 (−0.10)93.21 (−0.17)
75%94.59 (−0.63)98.36 (−0.03)96.61 (+0.08)94.34 (−0.50)
spambase25%80.60 (+0.08)87.08 (+0.11)81.31 (+0.13)79.04 (+0.01)
50%86.40 (+0.33)92.57 (+0.15)88.75 (+0.26)85.39 (+0.26)
75%89.92 (+0.02)95.16 (+0.14)92.42 (+0.25)89.26 (−0.01)
a8a25%78.48 (+0.09)79.66 (−0.46)92.03 (−0.31)66.05 (−1.03)
50%81.27 (+0.12)85.06 (−0.39)94.46 (−0.21)72.18 (−1.26)
75%82.95 (+0.02)87.59 (−0.37)95.56 (−0.19)75.52 (−1.13)
spam25%90.93 (−0.06)95.90 (−0.14)92.21 (−0.30)89.34 (−0.04)
50%94.04 (−0.36)98.06 (−0.10)95.72 (−0.15)93.28 (−0.45)
75%95.15 (−0.46)98.59 (−0.11)96.94 (+0.00)94.57 (−0.59)

These results indicate that the unified configuration yields performance comparable to the original dataset-specific configurations across the four datasets.

Q1: Learning Rate and Network Depth Selection

We performed hyperparameter optimization at a feature observation probability of p=0.5p=0.5 and directly applied the selected configurations to p=0.25p=0.25 and p=0.75p=0.75. We used Balanced Accuracy (BA) as the hyperparameter selection criterion. The specific configurations for individual datasets are reported in Appendix D.4, while the learning-rate sensitivity analysis is provided in Appendix F.1. Regarding network depth, we favored a lightweight architecture to balance model capacity and computational efficiency. Throughout the selection process, we prioritized configurations that performed consistently across different datasets.

Q2: OnlineNet vs. OnlineMLP

Thank you for your question. We compared OnlineNet against a plain MLP variant, referred to as OnlineMLP, using the same causally normalized inputs and feature-availability masks. The results are reported below. Each cell shows the OnlineMLP score, followed by the difference (OnlineMLP − OnlineNet) in percentage points.

DatasetAccuracyAUROCAUPRCBalanced Accuracy
wdbc78.56 (-4.32)86.41 (-2.88)80.85 (-3.45)74.29 (-6.42)
spam86.16 (-4.83)93.20 (-2.84)88.34 (-4.17)82.38 (-7.00)
diabetes_f65.76 (-1.25)60.56 (-3.58)43.36 (-6.01)52.32 (-3.64)
spambase77.85 (-2.67)83.98 (-2.99)77.39 (-3.79)75.36 (-3.67)
LUdata63.07 (-3.28)70.78 (-3.26)70.83 (-3.34)62.89 (-3.41)

Q3: Random Seeds

Different random seeds affect only model parameter initialization. As specified in Section 5.1 (Implementation Details), we use five fixed random seeds 0,1,2,3,4\\{0,1,2,3,4\\} for stochastic methods on real-world datasets and three seeds 0,1,2\\{0,1,2\\} on synthetic datasets. We report the mean and standard deviation to assess run-to-run variability. The explicit seed settings facilitate reproducibility. OLVF and OLIFL are deterministic and therefore evaluated only once, as their results are seed-invariant.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.