Concept Drift Detection via Equal Intensity k-Means Space Partitioning.

Area of Science:

• Machine Learning
• Data Mining
• Statistical Learning

Background:

• Data streams present challenges for statistical classification due to evolving data distributions, known as concept drift.
• Existing histogram-based drift detection methods often use arbitrary space partitioning, leading to potential blind spots and reduced accuracy.
• Unsupervised methods for improving drift detection accuracy in histogram-based approaches are needed.

Purpose of the Study:

• To propose a novel cluster-based histogram method, equal intensity k-means space partitioning (EI-kMeans), for enhanced concept drift detection.
• To improve the sensitivity of drift detection by minimizing partition creation in distribution offset regions.
• To dynamically determine histogram bin shapes and numbers for optimal drift detection.

Main Methods:

• Developed EI-kMeans, a cluster-based histogram approach for unsupervised concept drift detection.
• Introduced a heuristic method to enhance drift detection sensitivity.
• Employed Pearson's chi-square test for distribution-independent statistical hypothesis testing.
• Designed three algorithms: greedy centroids initialization, cluster amplify-shrink, and a drift detection algorithm.

Main Results:

• EI-kMeans demonstrated advantages over existing methods in concept drift detection.
• Experiments on synthetic and real-world datasets confirmed the efficacy of EI-kMeans.
• The dynamic determination of bins and shapes improved drift detection capabilities.

Conclusions:

• EI-kMeans offers a more accurate and sensitive approach to concept drift detection in streaming data.
• The proposed method addresses limitations of traditional grid-based and tree-based histogram partitioning.
• Retraining learners upon drift detection is recommended for adaptation.