Query by transduction

Area of Science:

• Machine Learning
• Statistical Learning Theory
• Data Mining

Background:

• Growing interest in transductive inference for machine learning applications.
• Active learning is crucial for efficient data labeling in stream-based environments.
• Existing active learning methods may not fully leverage transductive principles.

Purpose of the Study:

• To extend transductive inference to active learning within a stream-based setting.
• To propose and evaluate a novel active learning algorithm named Query-by-Transduction (QBT).

Main Methods:

• Introduced Query-by-Transduction (QBT), an active learning algorithm utilizing p-values from transduction.
• Established theoretical connections between QBT, Query-by-Committee (QBC), Bayesian testing, KL-divergence, and Shannon information.
• Empirically validated QBT on binary and multi-class classification tasks using Support Vector Machines (SVM).

Main Results:

• Demonstrated the feasibility and utility of QBT across various classification problems.
• QBT achieved favorable mean generalization performance compared to random sampling and other active learning strategies.
• Experimental results confirmed QBT's effectiveness in the stream-based active learning context.

Conclusions:

• QBT offers a novel and effective approach to active learning in stream-based scenarios.
• The algorithm leverages transduction principles to efficiently query informative labels.
• QBT shows competitive or superior performance against established active learning methods.