Low-rank robust online distance/similarity learning based on the rescaled hinge loss

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Online metric learning algorithms face scalability challenges with large datasets and high dimensions.
• Existing Passive/Aggressive (PA) methods, while fast, lack robustness against outliers and label noise due to Hinge loss.

Purpose of the Study:

• To develop robust and scalable online metric learning algorithms.
• To address limitations of existing methods concerning data noise and computational efficiency.

Main Methods:

• Formulated online Distance/Similarity learning with a robust Rescaled Hinge loss function.
• Proposed low-rank online methods to learn projection matrices, enhancing scalability to data dimension.
• Developed an efficient one-pass triplet construction algorithm to overcome limitations of batch sampling.

Main Results:

• The proposed Rescaled Hinge loss enhances robustness against outliers and label noise.
• Low-rank approaches reduce computational cost while preserving metric discrimination power.
• The novel triplet construction method is efficient and effective, outperforming batch sampling.

Conclusions:

• The developed robust and scalable online metric learning methods significantly outperform state-of-the-art techniques.
• These advancements are crucial for applications dealing with noisy data and large-scale datasets.