Improved Graph Embedding for Robust Recognition with outliers.

Area of Science:

• Biomedical Signal Processing
• Machine Learning
• Data Science

Background:

• Artifacts in biomedical signals (gene expression, sonar, EEG) disrupt data structure and bias conventional dimension reduction methods like L2 norm-based graph embedding (GE).
• Existing GE methods are sensitive to outliers, leading to inaccurate sub-structural feature extraction.

Purpose of the Study:

• To develop a robust graph embedding method less susceptible to artifacts and outliers in biomedical data.
• To enhance the accuracy and reliability of dimension reduction for contaminated datasets.

Main Methods:

• Defined a novel graph embedding (GE) model within the L1 norm space.
• Employed a maximization strategy to solve the L1 norm-based GE model, effectively mitigating outlier influence.
• Evaluated the method's performance against conventional GE techniques under various outlier conditions.

Main Results:

• The L1 norm-based GE demonstrated superior stability in estimating hyperplanes compared to conventional L2 norm-based methods.
• The proposed L1 GE method exhibited increased robustness to outliers, resulting in higher classification accuracy across diverse datasets.
• Quantitative evaluations confirmed the effectiveness of L1 GE in handling datasets with significant artifact contamination.

Conclusions:

• The L1 norm-based graph embedding offers a more reliable approach for dimension reduction in biomedical signal processing.
• This method is particularly beneficial for datasets contaminated with outliers, improving the extraction of meaningful mapping information.
• The proposed L1 GE provides a robust alternative for analyzing noisy biomedical data, enhancing downstream machine learning tasks.