AIGS: Adaptive Incremental Gating System for Online Representation Learning in Non-Stationary Data Streams

SiRui He; Kai Liang Lew; Chui Zi Ong; Chean Khim Toa

cs.LGOct 2, 2026

AIGS: Adaptive Incremental Gating System for Online Representation Learning in Non-Stationary Data Streams

Authors: SiRui He, Kai Liang Lew, Chui Zi Ong, Chean Khim Toa

Organizations: Xiamen University Malaysia Sepang, Malaysia · Faculty of Engineering & Technology, Multimedia University Melaka, Melaka, Malaysia · School of Economics and Management, Xiamen University Malaysia Sepang, Malaysia · School of Computing and Data Science, Xiamen University Malaysia Sepang, Malaysia

Abstract

Real-time data streams in Web of Things (WoT) and edge computing environments often evolve through latent regime changes. For online representation learning under strict computational constraints, the central problem is resolving the stability-plasticity dilemma: keeping useful historical knowledge while rapidly reacting to concept drift. Existing methods employ fixed update schedules or rolling windows. However, they suffer from parameter ossification during sudden shifts and waste computational resources when the stream remains stable. This paper proposes the Adaptive Incremental Gating System (AIGS), a lightweight closed-loop state-aware adaptation framework. AIGS introduces the Shock Ratio, an endogenous residual feedback mechanism that normalizes current reconstruction error against recent variation. This signal drives a Continuous Plasticity Controller that smoothly interpolates between learning plasticity and memory retention. By treating representation learning as a closed-loop control mechanism, AIGS avoids catastrophic forgetting and maintains a strictly linear $\mathcal{O}\left(k\cdot d\right)$ per-step complexity suitable for latency-sensitive edge devices. Experiments on real-world smart city dynamic streams-spanning traffic networks, meteorological systems, and industrial infrastructure-demonstrate distinct domain-dependent advantages. On Electricity Transformer Temperature datasets, AIGS achieves preventative early-warning lead times of 8.31 (ETTm1) and 9.88 (ETTm2) steps under gradual degradation. On Performance Measurement System traffic datasets, it shows significantly faster post-shift recovery after abrupt mutations. On the highly noisy Weather dataset, it improves anomaly recall while resisting stochastic noise overfitting. These findings establish AIGS as a practical, plug-and-play adapter for resource-constrained edge monitoring systems.

Figures & tables

Figure 1. Adaptive Incremental Gating System (AIGS) Architecture Architecture diagram showing the AIGS stream input, centering, subspace projection, residual sensing, Shock Ratio, soft gate, adaptive update, and downstream output.

Datasets	Dimension	Time Interval	Timesteps
PEMS04	307	5 mins	16,992
PEMS08	170	5 mins	17,856
ETTm1	7	15 mins	69,680
ETTm2	7	15 mins	69,680
Weather	14	10 mins	52,259

Table 1. Summary of datasets

	PEMS04						PEMS08
	Short-term (3 6)			Long-term (12 24)			Short-term (3 6)			Long-term (12 24)
Model	RMSE	MAE	ART	RMSE	MAE	ART	RMSE	MAE	ART	RMSE	MAE	ART
AIGS+GRU	$\underline{30.56\pm 0.07}$	$\underline{\underline{18.86\pm 0.50}}$	$\mathbf{12.76\pm 1.44}$	$\mathbf{34.35\pm 3.90}$	$\mathbf{23.38\pm 3.93}$	$\mathbf{17.38\pm 2.44}$	$\underline{28.65\pm 0.26}$	$\underline{\underline{18.47\pm 0.33}}$	$\mathbf{10.96\pm 3.28}$	$\mathbf{34.98\pm 6.39}$	$\mathbf{24.81\pm 5.97}$	$\mathbf{16.42\pm 0.59}$
T-GCN	$48.29\pm 2.04$	$36.68\pm 1.76$	$28.40\pm 3.39$	$63.09\pm 9.30$	$49.20\pm 7.45$	$30.96\pm 3.98$	$47.83\pm 1.79$	$36.38\pm 1.23$	$32.37\pm 0.10$	$59.25\pm 7.55$	$45.42\pm 6.01$	$27.80\pm 6.44$
Naked_GRU	$\underline{\underline{27.49\pm 1.04}}$	$\underline{22.06\pm 1.02}$	$\underline{19.56\pm 0.23}$	$52.36\pm 15.33$	$41.40\pm 11.55$	$30.41\pm 7.24$	$\underline{\underline{26.76\pm 3.23}}$	$\underline{19.46\pm 2.52}$	$17.91\pm 0.88$	$43.46\pm 11.40$	$32.73\pm 9.50$	$31.07\pm 6.20$
iTransformer	$32.75\pm 2.29$	$23.88\pm 1.88$	$30.33\pm 6.36$	$\underline{48.89\pm 10.72}$	$\underline{36.96\pm 8.50}$	$33.90\pm 0.80$	$29.01\pm 2.02$	$20.91\pm 1.66$	$22.44\pm 4.37$	$\underline{43.18\pm 9.23}$	$\underline{32.64\pm 7.38}$	$27.22\pm 2.38$

Table 2. PEMS04 and PEMS08 Experimental Results

Model	Short-term (8 16 32)			Mid-term (48 96 192)			Long-term (288 672 1344)
Model	RMSE	MAE	Avg_Lead	RMSE	MAE	Avg_Lead	RMSE	MAE	Avg_Lead
AIGS+GRU	$\underline{\underline{1.49\pm 0.15}}$	$\underline{\underline{1.06\pm 0.10}}$	$\mathbf{8.31\pm 2.95}$	$\underline{1.60\pm 0.08}$	$\underline{\underline{1.12\pm 0.09}}$	$\mathbf{8.09\pm 0.92}$	$\mathbf{1.65\pm 0.01}$	$\mathbf{1.18\pm 0.02}$	$\mathbf{7.14\pm 0.87}$
ADWIN+GRU	$1.66\pm 0.04$	$1.20\pm 0.03$	$-2.55\pm 0.39$	$1.67\pm 0.03$	$1.19\pm 0.02$	$-3.74\pm 0.94$	$1.70\pm 0.01$	$\underline{1.22\pm 0.01}$	$-0.64\pm 0.91$
CEP	$1.66\pm 0.03$	$1.17\pm 0.05$	$-3.71\pm 2.48$	$1.62\pm 0.08$	$\underline{1.13\pm 0.08}$	$-3.56\pm 1.02$	$\underline{1.68\pm 0.01}$	$\underline{\underline{1.19\pm 0.02}}$	$-1.00\pm 1.26$
DLinear	$1.58\pm 0.10$	$1.15\pm 0.09$	$-3.29\pm 1.18$	$\mathbf{1.55\pm 0.07}$	$\mathbf{1.08\pm 0.06}$	$-3.90\pm 1.28$	$1.70\pm 0.01$	$1.24\pm 0.02$	$\underline{0.51\pm 5.70}$
FSNet	$1.56\pm 0.13$	$1.12\pm 0.10$	$-2.23\pm 1.41$	$1.62\pm 0.07$	$1.14\pm 0.08$	$-4.51\pm 0.49$	$\underline{1.68\pm 0.01}$	$\underline{\underline{1.19\pm 0.02}}$	$-1.33\pm 1.73$

Table 3. Experimental Results on ETTm1 Dataset

Model	Short-term (8 16 32)			Mid-term (48 96 192)			Long-term (288 672 1344)
Model	RMSE	MAE	Avg_Lead	RMSE	MAE	Avg_Lead	RMSE	MAE	Avg_Lead
AIGS+GRU	$\underline{1.90\pm 0.30}$	$\underline{1.45\pm 0.24}$	$\underline{\underline{5.87\pm 1.15}}$	$\mathbf{2.21\pm 0.11}$	$\underline{\underline{1.65\pm 0.10}}$	$\mathbf{9.88\pm 0.96}$	$\mathbf{2.28\pm 0.02}$	$\mathbf{1.70\pm 0.02}$	$\mathbf{8.22\pm 0.86}$
ADWIN+GRU	$1.91\pm 0.32$	$1.48\pm 0.26$	$4.09\pm 2.68$	$2.31\pm 0.08$	$1.73\pm 0.08$	$0.08\pm 1.21$	$\underline{2.34\pm 0.04}$	$\underline{1.76\pm 0.03}$	$1.68\pm 3.53$
CEP	$1.91\pm 0.32$	$1.47\pm 0.25$	$3.97\pm 2.22$	$\underline{2.26\pm 0.11}$	$\underline{1.70\pm 0.11}$	$-0.27\pm 1.55$	$\underline{2.34\pm 0.04}$	$\underline{1.76\pm 0.03}$	$1.62\pm 3.45$
DLinear	$\underline{\underline{1.82\pm 0.21}}$	$\underline{\underline{1.37\pm 0.15}}$	$5.43\pm 2.65$	$\underline{\underline{2.23\pm 0.12}}$	$\mathbf{1.63\pm 0.10}$	$0.14\pm 1.65$	$2.63\pm 0.10$	$1.93\pm 0.07$	$0.69\pm 1.89$
FSNet	$1.92\pm 0.31$	$1.47\pm 0.24$	$3.10\pm 1.19$	$\underline{2.27\pm 0.07}$	$\underline{1.70\pm 0.07}$	$2.21\pm 0.82$	$\underline{\underline{2.30\pm 0.03}}$	$\underline{\underline{1.73\pm 0.02}}$	$1.95\pm 3.67$

Table 4. Experimental Results on ETTm2 Dataset

Model	Short-term (36 72)		Mid-term (144 288 432)		Long-term (1008 1440)
Model	RMSE	MAE	RMSE	MAE	RMSE	MAE
AIGS+GRU	$\underline{0.27\pm 0.01}$	$0.21\pm 0.01$	$\underline{0.27\pm 0.00}$	$\underline{0.21\pm 0.00}$	$\underline{\underline{0.27\pm 0.00}}$	$\underline{\underline{0.21\pm 0.00}}$
ADWIN+GRU	$\underline{0.27\pm 0.00}$	$\underline{0.20\pm 0.00}$	$\mathbf{0.25\pm 0.01}$	$\mathbf{0.19\pm 0.00}$	$\mathbf{0.26\pm 0.01}$	$\mathbf{0.20\pm 0.00}$
CEP	$\underline{0.27\pm 0.01}$	$\underline{0.20\pm 0.01}$	$\underline{\underline{0.26\pm 0.01}}$	$\underline{\underline{0.20\pm 0.01}}$	$\underline{0.28\pm 0.01}$	$\underline{0.22\pm 0.01}$
DLinear	$\underline{\underline{0.26\pm 0.00}}$	$\underline{\underline{0.19\pm 0.00}}$	$0.29\pm 0.02$	$0.22\pm 0.02$	$0.32\pm 0.00$	$0.25\pm 0.00$
FSNet	$\underline{\underline{0.26\pm 0.00}}$	$0.21\pm 0.01$	$\underline{\underline{0.26\pm 0.01}}$	$\underline{\underline{0.20\pm 0.01}}$	$\underline{\underline{0.27\pm 0.00}}$	$\underline{\underline{0.21\pm 0.00}}$

Table 5. Experimental Results on Weather Dataset

Figure 2. Global Anomaly Recall and False Alarm Rate (FAR) Across Different Prediction Horizons. Plot comparing anomaly recall and false alarm rate across prediction horizons under the rolling 3-sigma event protocol.

Figure 3. Analysis of the Trade-off between Inference Time (s) and RMSE Score. Lollipop plot comparing inference time and RMSE for evaluated models to show the efficiency and accuracy trade-off.

Figure 4. Weather Radar Performance for Recall and 100-FAR (%) across Different Horizons. Radar chart comparing Weather recall and 100-FAR across horizons for AIGS and baseline models.

Model	Time(s)	Peak RAM	Avg RAM	Peak VRAM	Avg VRAM
Naked_GRU	118.30	5.1525	2.5561	160.5005	65.7165
Rolling_PCA+GRU	133.05	4.8944	2.2919	159.8794	65.3334
ReVIN+GRU	144.61	5.2061	2.6017	161.2690	65.3775
AIGS+GRU	130.94	5.7604	2.7656	160.0381	65.8103

Table 6. System Resource Utilization and Efficiency Analysis

Model	Recall $\uparrow$	$AvgLead$ $\uparrow$	RMSE $\downarrow$
AIGS_CCIPCA+GRU	98.8846	11.3024	0.2897
CCIPCA+GRU	55.2769	8.1009	0.2926
AIGS_OPAST+GRU	8 9.1129	10.6352	0.2898
OPAST+GRU	53.7544	7.2879	0.2789
AIGS_LinearAE+GRU	93.1284	8.1066	0.2748
LinearAE+GRU	54.5897	7.0291	0.2792

Table 7. Performance on Different Integration

Model	ART $\downarrow$	$AvgLead$ $\uparrow$	Time $\downarrow$	MAE $\downarrow$	$R^{2}$ $\uparrow$	RMSE $\downarrow$	Recall $\uparrow$
AIGS(Inc.)+GRU	9.4662	8.3106	16.1222	1.0704	0.2528	1.5068	96.8254
AIGS+GRU (Ours)	9.4115	8.3106	15.3366	1.0540	0.2707	1.4886	96.8254
GHA+GRU	10.6264	-2.8391	12.7210	1.1355	0.1623	1.6015	64.7481
Naked_GRU	9.7900	-0.9356	16.6441	1.1034	0.2252	1.5359	75.9019
Rolling_PCA+GRU	10.1623	-2.5119	18.5587	1.1078	0.2067	1.5545	72.7273

Table 8. Ablation Study Results

Figure 5. Hyperparameter Optimization for Sensitivity of RMSE and Switching Threshold (

T

) vs Magnification (

M

). Bivariate surface showing RMSE variation across switching threshold and magnification settings.

Figure 6. Global Sensitivity Analysis of Stability vs. Alertness across Prediction Horizons ( $H$ =8 to $H$ =1344). Heatmap showing Sobol sensitivity scores for stability and alertness metrics across prediction horizons.

Time Series Representation Learning

Time series representation learning develops compact, transferable features that capture temporal patterns for forecasting, classification, retrieval, and anomaly detection.

See all 94 papers

Oct 23, 2026

Online Continual Learning in Edge-Cloud Applications

Deployed machine learning models in real-world IoT and edge/cloud systems face a critical reliability bottleneck: the requirement for online adaptation to non-stationary data streams while operating without access to the full original training dataset. Performance is severely degraded when models encounter Out-of-Distribution (OOD) inputs or suffer from catastrophic forgetting. Addressing these challenges requires adaptation mechanisms capable of learning from evolving data streams while preserving previously acquired knowledge. This paper proposes an edge-cloud methodology for integrating Continual Learning (CL) with Out-of-Distribution (OOD) Detection and Concept Drift Detection in IoT applications. The methodology incorporates ODIN and ECDD as alternative mechanisms for identifying informative samples for annotation and triggering rehearsal-based model updates. Experimental results across automotive, healthcare, and agricultural datasets demonstrate the applicability of the proposed methodology and its potential to support model adaptation while mitigating catastrophic forgetting.

Charalampos Davalas, Dimitrios Michail, Christos Diou +2

Department of Informatics and Telematics, Harokopio University of Athens, Athens, 17778, Greece · School of Electrical and Computer Engineering, National Technical University of Athens, Athens, 15773, Greece

Online Continual Learning Edge ML Internet of Things

Ask PDF

Feb 27, 2026

FastODT: A tree-based framework for efficient continual learning

Machine learning models deployed in real-world settings must operate under evolving data distributions and constrained computational resources. This challenge is particularly acute in non-stationary domains such as energy time series, weather monitoring, and environmental sensing. To remain effective, models must support adaptability, continuous learning, and long-term knowledge retention. This paper introduces a oblivious tree-based model with Hoeffding bound controlling its growth. It seamlessly integrates rapid learning and inference with efficient memory management and robust knowledge preservation, thus allowing for online learning. Extensive experiments across energy and environmental sensing time-series benchmarks demonstrate that the proposed framework achieves performance competitive with, and in several cases surpassing, existing online and batch learning methods, while maintaining superior computational efficiency. Collectively, these results demonstrate that the proposed approach fulfills the core objectives of adaptability, continual updating, and efficient retraining without full model retraining. The framework provides a scalable and resource-aware foundation for deployment in real-world non-stationary environments where resources are constrained and sustained adaptation is essential.

Daniel Bretsko, Piotr Walas, Devashish Khulbe +3

Masaryk University, Brno, Czech Republic · Warsaw University of Technology, Warsaw, Poland · Mycroft Mind, Brno, Czech Republic +2

Online Continual Learning Decision Tree Learning Memory-Efficient Training

Ask PDF

Mar 8, 2026

Gated Adaptation for Continual Learning in Human Activity Recognition

Wearable sensors in Internet of Things (IoT) ecosystems increasingly support applications such as remote health monitoring, elderly care, and smart home automation, all of which rely on robust human activity recognition (HAR). Continual learning systems must balance plasticity (learning new tasks) with stability (retaining prior knowledge), yet AI models often exhibit catastrophic forgetting, where learning new tasks degrades performance on earlier ones. This challenge is especially acute in domain-incremental HAR, where on-device models must adapt to new subjects with distinct movement patterns while maintaining accuracy on prior subjects without transmitting sensitive data to the cloud. We propose a parameter-efficient continual learning framework based on channel-wise gated modulation of frozen pretrained representations. Our key insight is that adaptation should operate through feature selection rather than feature generation: by restricting learned transformations to diagonal scaling of existing features, we preserve the geometry of pretrained representations while enabling subject-specific modulation. We provide a theoretical analysis showing that gating implements a bounded diagonal operator that limits representational drift compared to unconstrained linear transformations. Empirically, freezing the backbone substantially reduces forgetting, and lightweight gates restore lost adaptation capacity, achieving stability and plasticity simultaneously. On PAMAP2 with 8 sequential subjects, our approach reduces forgetting from 39.7% to 16.2% and improves final accuracy from 56.7% to 77.7%, while training less than 2% of parameters. Our method matches or exceeds standard continual learning baselines without replay buffers or task-specific regularization, confirming that structured diagonal operators are effective and efficient under distribution shift.

Reza Rahimi Azghan, Gautham Krishna Gudur, Mohit Malu +4

Arizona State University, Phoenix, AZ, USA · The University of Texas at Austin, Austin, TX, USA

Domain-Incremental Learning Human Activity Recognition Parameter-Efficient Fine-Tuning

Ask PDF

Abstract

Time Series Representation Learning

Explore similar work

Online Continual Learning in Edge-Cloud Applications

FastODT: A tree-based framework for efficient continual learning

Gated Adaptation for Continual Learning in Human Activity Recognition