Hybrid Active-Online Learning Framework for Label-Efficient Concept Drift Adaptation in Optical Network Failure Detection

Technical Summary: Hybrid Active-Online Learning for Optical Network Failure Detection

Problem Statement
Effective failure management in optical networks is critical for maintaining continuous service. While machine learning (ML) methods have been proposed to automate failure detection, static models often fail when the underlying data distribution changes—a phenomenon known as Concept Drift (CD). In optical networks, CD arises from factors such as equipment aging and varying channel properties. A specific challenge occurs when models trained on "soft failure" data encounter rare "hard failure" events, which are difficult to simulate and for which labeled data is scarce.

Traditional online learning adapts to CD by updating models sequentially on streaming data. However, standard online approaches assume that labels are constantly available for every sample. In real-world deployments, this assumption rarely holds, as expert verification or automated oracles are not always accessible. Consequently, there is a need for a method that adapts to concept drift while operating under severe label scarcity.

Methodology
The authors propose a Hybrid Active-Online Learning Framework that integrates the adaptability of online learning with the label-efficiency of active learning.

• Dataset and Features: The study utilizes a dataset generated from an experimental testbed containing four features: Bit-Error-Rate (BER) and Optical Signal-to-Noise Ratio (OSNR) measured at both the transmitter and receiver. The data is split into Soft Failure (SFD) and Hard Failure (HFD) sets.
• System Architecture: Three models were compared:
  1. Static Model: Pre-trained on SFD and used for inference only on new samples.
  2. Supervised Online Model: Pre-trained on SFD, then updated in a prequential manner (predict, then learn) using every available sample and its label.
  3. Hybrid (Active+Online) Model: Pre-trained on SFD, then processes streaming data by first predicting. It employs Uncertainty Sampling to calculate the margin (difference in probability) between failure and normal classes.
• Labeling Strategy: The hybrid system queries a label only if the uncertainty margin exceeds a tunable Margin Threshold and if the system has available "budget tokens." If queried, the model updates; otherwise, the sample is skipped to preserve the labeling budget. The budget refills incrementally.
• Drift Detection: A two-stage procedure (Page-Hinkley Test followed by Kolmogorov-Smirnov test) was deployed to identify concept drifts in the OSNR_Rx feature, which exhibited the highest correlation with failure labels. This detection was used for visualization and analysis but did not directly control the labeling budget.
• Algorithm: The Adaptive Random Forest (ARF) ensemble was selected for all three systems due to its suitability for online learning tasks.

Key Contributions
The primary contribution of this work is the first joint application of active and online learning for optical network failure detection. The framework aims to provide the "best of both worlds": the ability of online learning to adapt to concept drift and the label-efficiency of active learning to operate in label-scarce environments.

Results
Experiments were conducted using the HFD as streaming data, with synthetic failure samples added to the end of the stream to simulate increased failure events.

• Accuracy and AUC: The hybrid approach achieved performance comparable to the fully supervised online model. By tuning the Margin Threshold to 0.7, the system achieved a rolling accuracy of 0.77 and an AUC of 0.59 in the final window. Crucially, this performance was attained while querying only 3.4% of the streaming samples (193 queries out of the original HFD stream).
• Adaptation to Drift: Both online models (supervised and hybrid) significantly outperformed the static model when encountering concept drift and hard failures. The hybrid model's accuracy improved as it learned from the few queried failure samples, eventually matching and briefly surpassing the supervised model's performance when synthetic samples were introduced.
• Latency: The hybrid system demonstrated negligible latency overhead. The mean per-event latency was 0.1650 ms, nearly identical to the static model (0.1624 ms). The online overhead for the hybrid model was only 0.0027 ms, a drastic reduction compared to the 0.3522 ms overhead of the fully supervised online model.

Significance and Claims
The paper claims that the proposed hybrid framework offers a practical solution for real-time monitoring in optical networks where labeling is constrained. By achieving near-ceiling accuracy and AUC scores with minimal labeling (3.4% of samples) and maintaining latency levels comparable to static inference, the method addresses the critical trade-off between adaptation performance and operational cost. The authors conclude that this approach enables effective concept drift adaptation without the prohibitive cost of continuous human or oracle labeling.