Unsupervised active learning based on hierarchical graph-theoretic clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Supervised active learning struggles with semantic gaps and new data categories.
• Existing methods lack adaptability to evolving label interpretations.
• Inefficiency in supervised active learning hinders accurate classification.

Purpose of the Study:

• To propose an unsupervised active learning framework to overcome supervised limitations.
• To enhance sample selection for new and evolving categories.
• To improve classification accuracy and reduce manual workload.

Main Methods:

• A hierarchical graph-theoretic clustering framework combining dominant-set and spectral clustering.
• Unsupervised active learning approach for efficient sample selection.
• Evaluation on network intrusion detection, image, and video classification datasets.

Main Results:

• The proposed framework effectively reduces manual classification workload.
• Maintains high accuracy in automatic classification tasks.
• Outperforms Support Vector Machine-based supervised active learning, especially with novel categories.

Conclusions:

• The unsupervised active learning framework offers an efficient and adaptable solution.
• Hierarchical graph-theoretic clustering addresses limitations of supervised active learning.
• Demonstrates superior performance in handling new data categories and classification challenges.