Active Learning With Multiple Kernels.

Area of Science:

• Machine Learning
• Kernel Methods
• Data Science

Background:

• Online Multiple Kernel Learning (OMKL) excels in nonlinear function learning.
• Random Feature (RF) approximation mitigates OMKL's curse of dimensionality.
• Label acquisition is often costly and time-consuming in real-world applications.

Purpose of the Study:

• Introduce stream-based active Multiple Kernel Learning (AMKL) for efficient learning with limited labeled data.
• Develop an adaptive kernel selection mechanism (AMKL-AKS) to enhance learning efficiency and accuracy.
• Theoretically prove AMKL's optimal sublinear regret and empirically validate AMKL-AKS's superiority.

Main Methods:

• Leveraging random feature approximation for OMKL.
• Implementing a novel selection criterion for data labeling in a streaming setting.
• Developing an adaptive kernel selection strategy to dynamically exclude irrelevant kernels.

Main Results:

• The proposed AMKL achieves an optimal sublinear regret of O(√T) with minimal labeled data.
• AMKL-AKS demonstrates superior performance compared to traditional OMKL, requiring fewer labels for similar accuracy.
• Numerical tests on real datasets confirm the effectiveness and efficiency of AMKL-AKS.

Conclusions:

• Stream-based active MKL offers an efficient solution for nonlinear function learning with costly labeling.
• AMKL-AKS enhances active learning efficiency and function learning accuracy by adaptively managing kernel dictionaries.
• The proposed methods provide a practical approach for real-world applications demanding cost-effective data labeling.