Learning With Asymmetric Kernels: Least Squares and Feature Interpretation.

Area of Science:

• Machine Learning
• Kernel Methods
• Data Science

Background:

• Asymmetric kernels are prevalent in real-world applications like directed graphs and conditional probability.
• Existing kernel-based learning methods predominantly require symmetric kernels, limiting the application of asymmetric ones.
• This limitation hinders the direct utilization of valuable asymmetric information in various learning tasks.

Purpose of the Study:

• To develop a novel classification method capable of directly utilizing asymmetric kernels.
• To extend the framework of least squares support vector machines (LS-SVM) to accommodate asymmetric kernels.
• To demonstrate the efficacy of the proposed method in scenarios where asymmetric information is critical.

Main Methods:

• Introduction of Asymmetric Kernel Least Squares (AsK-LS), a new classification method.
• Leveraging the kernel trick with both source and target features, even when unknown.
• Ensuring computational efficiency comparable to methods using symmetric kernels.

Main Results:

• AsK-LS successfully learns from asymmetric features (source and target).
• The kernel trick remains applicable within the AsK-LS framework.
• Experimental results across diverse datasets (Corel, PASCAL VOC, Satellite, directed graphs, UCI) show superior performance compared to symmetrization-based methods.
• AsK-LS demonstrates significant advantages when asymmetric information is crucial.

Conclusions:

• AsK-LS is the first classification method that directly utilizes asymmetric kernels.
• The method offers a computationally efficient and effective solution for learning with asymmetric data.
• AsK-LS significantly advances kernel-based learning by enabling the direct use of asymmetric kernel properties.