Active Learning Based on Locally Linear Reconstruction.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Active learning aims to select the most representative data points for model training.
• Optimum Experimental Design (OED) is a common active learning technique, focusing on global Euclidean structure.
• Existing OED methods often overlook the local manifold structure of data.

Purpose of the Study:

• To propose a novel active learning algorithm that incorporates local data structure.
• To define representative points based on their ability to reconstruct the entire dataset using local information.
• To address limitations of existing methods that focus solely on global properties.

Main Methods:

• Developed a new active learning algorithm considering local manifold structure.
• Utilized Locally Linear Reconstruction (LLR), a transductive learning algorithm.
• Employed sequential and convex optimization schemes to solve the reconstruction problem.

Main Results:

• The proposed method effectively leverages local data structure for point selection.
• Locally Linear Reconstruction (LLR) demonstrated the ability to reconstruct data points using neighbor information.
• Experimental results validated the effectiveness of the novel active learning approach.

Conclusions:

• The novel active learning algorithm enhances data point selection by considering local geometry.
• LLR provides a robust framework for identifying representative points based on local reconstructions.
• This approach offers a promising alternative to traditional OED methods in active learning.