Absent Multiple Kernel Learning Algorithms.

Area of Science:

• Machine Learning
• Data Science

Background:

• Multiple Kernel Learning (MKL) optimally combines sample features for classification.
• Existing MKL methods struggle with missing data channels, a common issue in real-world applications.

Purpose of the Study:

• To propose novel algorithms that effectively handle missing data in MKL.
• To develop methods that classify samples based on observed channels without imputation.

Main Methods:

• Introduction of three Absent Multiple Kernel Learning (AMKL) algorithms.
• Directly classifying samples using observed channels, bypassing imputation.
• Defining sample-specific margins in relevant spaces and maximizing the minimum margin.
• Developing two-step iterative algorithms and a convex reformulation using the representer theorem.

Main Results:

• AMKL algorithms demonstrate superior performance compared to imputation-based methods.
• Performance improvement increases with a higher missing data ratio.
• Theoretical justification provided through a generalization error bound.

Conclusions:

• The proposed AMKL algorithms offer a robust solution for MKL with missing data.
• These methods provide a significant advancement over traditional imputation approaches.
• AMKL algorithms are effective and theoretically sound for practical applications with incomplete datasets.