Online Learning With Adaptive Rebalancing in Nonstationary Environments.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Handling large, sequential data streams presents challenges in machine learning.
• Nonstationary environments and class imbalance significantly complicate online learning tasks.
• Existing methods often struggle with the dual challenge of concept drift and imbalanced data.

Purpose of the Study:

• To address the largely unexplored area of online learning from nonstationary and imbalanced data.
• To introduce a novel algorithm, Adaptive REBAlancing (AREBA), designed for these challenging data conditions.
• To provide new insights into maintaining class balance during online learning with concept drift.

Main Methods:

• Proposed the Adaptive REBAlancing (AREBA) algorithm.
• AREBA selectively includes majority and minority examples from the data stream.
• An adaptive mechanism within AREBA continually maintains class balance among selected examples.

Main Results:

• AREBA demonstrated significant improvements in both learning speed and learning quality compared to strong baselines.
• Extensive experiments on synthetic and real-world data confirmed AREBA's effectiveness across various imbalance rates and concept drift types.
• The algorithm consistently outperformed other state-of-the-art methods.

Conclusions:

• The proposed AREBA algorithm offers a robust solution for online learning in nonstationary and imbalanced environments.
• AREBA's adaptive balancing mechanism is key to its superior performance.
• Public code availability facilitates further research and application in this domain.