PB-OEL: A Performance-Bounded Online Ensemble Learning Framework With Mixed Feedback for Real-Time Safety Assessment

2026-07-09 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Data Science & Analytics · Depth: Expert, extended

Summary

PB-OEL, a performance-bounded online ensemble learning method based on multi-armed bandits, addresses the challenge of theoretically guaranteeing ensemble classifier performance in online learning, where most weight allocation strategies are heuristic. It incorporates multi-armed bandits with expert advice to dynamically update base classifier weights and make predictions. A theoretical framework is established, demonstrating that the ensemble classifier's performance bound can exceed any base classifier's performance when the data stream length is sufficiently large. The method also derives performance bounds for scenarios with limited annotations. Extensive experiments on 13 benchmark datasets and a real-time safety assessment task dataset validate the theoretical bounds and show PB-OEL outperforms other leading methods, achieving an average accuracy of 0.893, surpassing the second-best method, ARF, by nearly 3%. The method's time complexity per sample is O(N(Tp + Td + Tt + K) + 2K).

Key takeaway

For Machine Learning Engineers building online classification systems, PB-OEL provides a robust solution with theoretical performance guarantees. You should consider implementing this multi-armed bandit-based ensemble method, especially for non-stationary data streams and real-time safety assessments. Its proven ability to surpass individual base classifier accuracy and outperform other leading methods, even with limited annotations, makes it a strong candidate for critical applications requiring high reliability.

Key insights

PB-OEL uses multi-armed bandits with expert advice to provide theoretical performance bounds for online ensemble learning, outperforming individual base classifiers.

Principles

• Online ensemble performance can be theoretically bounded.
• MAB-EA optimizes base classifier weight allocation.
• Periodic weight resets enhance ultimate performance bounds.

Method

PB-OEL integrates MAB-EA to dynamically adjust base classifier weights using an REXP4 algorithm, predicting labels and updating weights based on rewards, with concept drift detection via Hoeffding inequality.

In practice

• Apply PB-OEL for real-time safety assessment.
• Use RVFL as diverse base classifiers.
• Configure alpha and N for performance/speed trade-offs.

Topics

• Online Ensemble Learning
• Multi-armed Bandits
• Performance Bounds
• Concept Drift Detection
• Real-time Safety Assessment
• REXP4 Algorithm

Code references

• THUFDD/THU-Concept-Drift-Datasets
• liuzy0708/Awesome_OL
• THUFDD/JiaolongDSMS_datasets

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Robotics Engineer

Related on AIssential

• Ensembles of Ensembles of Ensembles: A Guide to Stacking — Ensembling diverse models across multiple layers enhances predictive performance and robustness for tabular and time series data.
• Efficient Online Proportional Sampling with Applications to Smoothed Online Learning — A novel data structure enables efficient online proportional sampling in high-dimensional, dynamically evolving systems, achieving tight bounds.
• Simplex-Constrained Sparse Bagging: Transitioning from Uniform Priors to Sparse Posteriors in Ensemble Learning — SCSB compresses and calibrates bagging ensembles by optimizing sparse voting weights, improving efficiency and confidence.
• A Framework for Evaluating and Benchmarking Concept Drift Detection Methods — Inconsistent evaluation of concept drift detection methods is addressed by a new benchmarking framework.
• Scaling Learning-based AEB with Massive Unlabeled Data — Stabilized meta-feedback semi-supervised learning effectively scales AEB using massive unlabeled data, significantly improving safety metrics in real-world deployment.
• Ensemble Diversity Optimization for Subjective Supervision — EDO uses a prediction-space framework with a signed diversity regularizer to model human subjectivity and improve calibration in subjective NLP tasks.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.