A Bi-Criteria Active Learning Algorithm for Dynamic Data Streams.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Active learning (AL) efficiently builds training sets with minimal supervision by querying informative instances.
• Streaming data presents challenges like evolving distributions and sampling bias, especially with concept drift.
• Existing AL methods struggle to adapt to dynamic data streams and mitigate sampling bias effectively.

Purpose of the Study:

• To propose a novel bi-criteria active learning (BAL) approach for efficient learning from data streams.
• To address challenges of concept drift and sampling bias in evolving data distributions.
• To enhance model adaptation and ensure training sets represent the true underlying data distribution.

Main Methods:

• BAL utilizes two selection criteria: label uncertainty and density-based criterion.
• The label uncertainty criterion selects instances with ambiguous class membership.
• The density-based criterion curbs sampling bias by weighting samples according to the underlying data distribution.
• A Bayesian online learning approach combines online classification (logistic regression) and online clustering (Gaussian mixture models).

Main Results:

• The proposed BAL method demonstrates high performance on synthetic and real-world benchmarks.
• BAL effectively handles evolving data distributions and concept drift in streaming data.
• The method successfully reduces sampling bias compared to state-of-the-art AL techniques.

Conclusions:

• The bi-criteria active learning (BAL) approach offers a robust solution for learning from data streams.
• BAL enhances model accuracy and adaptability in dynamic environments by managing uncertainty and data distribution.
• This method provides a significant advancement in active learning for streaming data applications.