Batch Mode Active Sampling based on Marginal Probability Distribution Matching.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Active Learning (AL) efficiently uses labeled data for training machine learning models.
• Batch-mode AL selects multiple samples concurrently, improving efficiency over single-sample selection.
• Existing methods often focus on individual sample informativeness, potentially missing collective benefits.

Purpose of the Study:

• To propose a novel batch-mode active learning criterion for improved classifier generalization.
• To minimize the distribution shift between labeled and unlabeled data post-annotation.
• To address the computational complexity of optimal subset selection.

Main Methods:

• Formulated a novel objective to minimize distribution difference between labeled and unlabeled data.
• Transformed the NP-hard optimization problem into convex quadratic programming and linear programming problems.
• Evaluated the approach on UCI datasets, a biomedical image dataset, and the 20 Newsgroups dataset.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art batch-mode active learning techniques.
• Incorporating uncertainty improved performance in later active learning iterations.
• Transfer learning enhanced classifier performance in initial iterations.

Conclusions:

• The novel distribution-minimizing criterion offers an effective approach for batch-mode active learning.
• Optimization techniques provide efficient solutions for the complex selection problem.
• Extensions incorporating uncertainty and transfer learning offer further performance improvements in specific scenarios.