Active learning for classifying data streams with unknown number of classes.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Data stream classification is challenging due to infinite data size and dynamic distribution changes.
• Concept drift and concept evolution pose significant problems for traditional classification models.
• Labeling data in streams is expensive and often infeasible, necessitating efficient learning strategies.

Purpose of the Study:

• To propose a novel stream-based active learning algorithm (SAL) for data stream classification.
• To address concept drift and concept evolution in data streams.
• To develop an algorithm that queries only the most informative samples to reduce future error.

Main Methods:

• Developed a stream-based active learning algorithm (SAL) adaptable to dynamic stream changes.
• Incorporated strategies to handle sampling bias and query drifting or new class samples.
• Utilized non-parametric Bayesian models, specifically Dirichlet mixture models and the stick breaking process, for online learning.

Main Results:

• SAL effectively adapts classification models to concept drift and evolution.
• The algorithm prioritizes querying samples that are expected to reduce future error.
• Demonstrated superior classification performance compared to state-of-the-art methods on real-world benchmarks.

Conclusions:

• SAL is the first active learning algorithm to explicitly address concept drift and evolution in data streams.
• The proposed method achieves higher average and average class accuracy in dynamic environments.
• Non-parametric Bayesian models are well-suited for online learning in data streams with SAL.