Active Learning in Context-Driven Stream Mining With an Application to Image Mining.

Area of Science:

• Computer Science
• Medical Imaging
• Machine Learning

Background:

• Real-time processing of image streams with contextual metadata presents significant challenges for prediction accuracy.
• Online learning of classifier performance is crucial when image sources are unknown and true labels are costly to obtain.

Purpose of the Study:

• To develop an effective image stream mining method that addresses the challenges of real-time prediction and online learning.
• To model image stream mining as an active, online contextual experts problem, utilizing image context for classifier selection.

Main Methods:

• Feature extraction and preprocessing of images.
• Online classifier selection based on image context.
• Development of an active learning algorithm to minimize the need for costly true labels.

Main Results:

• The proposed active learning algorithm achieves sublinear regret.
• High accuracy in diagnosing breast cancer from cellular images was achieved by obtaining a small fraction of true labels.
• The method demonstrates potential for applications in video surveillance and traffic monitoring.

Conclusions:

• The developed image stream mining method effectively handles real-time data and learns classifier performance online.
• Active learning significantly reduces the cost of obtaining true labels while maintaining high diagnostic accuracy.
• The approach is applicable to various real-world scenarios requiring intelligent analysis of image streams.