Automatic Tuning of the RBF Kernel Parameter for Batch-Mode Active Learning Algorithms: A Scalable Framework.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Batch-mode active learning reduces data labeling costs by intelligently selecting samples for annotation.
• Radial-basis function (RBF) kernels are commonly used in active learning, but their parameter tuning is critical and often manual.
• Efficiently selecting valuable unlabeled samples is key to optimizing active learning performance.

Purpose of the Study:

• To propose a novel method for automatic kernel parameter tuning in batch-mode active learning.
• To develop frameworks that integrate automatic tuning into existing active learning algorithms.
• To enhance the efficiency and scalability of batch-mode active learning.

Main Methods:

• A hypothesis-margin-based criterion function is proposed for automatic kernel parameter tuning.
• Three frameworks are developed to incorporate this tuning function into batch-mode active learning algorithms.
• The frameworks allow for single-stage or multi-stage kernel parameter tuning, enabling coarse-to-fine sample importance evaluation.

Main Results:

• The proposed automatic tuning method effectively optimizes the RBF kernel parameter.
• The developed frameworks successfully integrate with existing batch-mode active learning algorithms.
• Experimental results demonstrate the feasibility and effectiveness of the proposed approach on large datasets.
• Scalability improvements were observed for algorithms with a decomposition property.

Conclusions:

• Automatic kernel parameter tuning is crucial for effective batch-mode active learning.
• The proposed hypothesis-margin-based criterion and integration frameworks offer a robust solution.
• The approach enhances sample selection accuracy and improves the scalability of active learning algorithms.