Cross-species Data Classification by Domain Adaptation via Discriminative Heterogeneous Maximum Mean Discrepancy

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Cross-species and cross-platform data classification face challenges due to differing feature spaces and statistical distributions.
• Traditional methods are inadequate for heterogeneous domain adaptation where feature spaces differ.
• Domain adaptation offers a strategy to address these classification challenges.

Purpose of the Study:

• To propose a novel heterogeneous domain adaptation approach for cross-species and cross-platform data classification.
• To address the challenge of reducing domain differences and correcting drift in heterogeneous domains.
• To leverage Maximum Mean Discrepancy (MMD) for measuring probability divergence in a common subspace.

Main Methods:

• Developed a Discriminative Heterogeneous Maximum Mean Discrepancy (DMMD) approach.
• Mapped data samples to a common low-dimensional subspace to minimize probability divergence.
• Employed a conjugate gradient algorithm on a Grassmann manifold to solve the DMMD model.
• Preserved known discriminative information during subspace mapping.

Main Results:

• The DMMD approach demonstrated superior performance compared to state-of-the-art methods on simulation and benchmark datasets.
• Experiments validated the effectiveness of DMMD for heterogeneous domain adaptation.
• Successful application to cross-platform and cross-species datasets confirmed its practical utility.

Conclusions:

• The proposed DMMD method is effective for heterogeneous domain adaptation in bioinformatics.
• Minimizing probability divergence in a common subspace is a viable strategy for cross-domain classification.
• DMMD offers a robust solution for classifying data across different species or platforms.