Robust Adaptive Embedded Label Propagation With Weight Learning for Inductive Classification.

Area of Science:

• Machine Learning
• Computer Science
• Data Science

Background:

• Semi-supervised learning is crucial for classification tasks with limited labeled data.
• Existing label propagation (LP) models often struggle with noisy features and separate weight optimization.
• The out-of-sample problem limits the generalizability of traditional LP methods.

Purpose of the Study:

• To develop a robust inductive semi-supervised label prediction model for classification.
• To integrate adaptive embedded label propagation with adaptive weight learning into a unified framework.
• To enhance classification performance by jointly optimizing representation and classification weights.

Main Methods:

• Proposed adaptive embedded label propagation with weight learning (AELP-WL) model.
• Jointly minimized reconstruction errors over embedded features and soft labels.
• Incorporated sparse decomposition to remove noise and learn weights over clean representations.
• Addressed the out-of-sample issue by including a regressive label approximation error.

Main Results:

• Achieved state-of-the-art results in classification tasks.
• Demonstrated improved performance by learning joint optimal weights for representation and classification.
• Effectively handled noisy data through sparse decomposition and adaptive weight learning.
• Naturally solved the out-of-sample problem via simultaneous minimization of errors.

Conclusions:

• AELP-WL offers a unified and robust framework for semi-supervised classification.
• The proposed method significantly enhances label estimation power and classification accuracy.
• AELP-WL provides a powerful solution for handling noisy data and out-of-sample predictions.