Training Dataset Curation by L₁-Norm Principal-Component Analysis for Support Vector Machines.

Area of Science:

• Machine Learning
• Data Science
• Computational Statistics

Background:

• Support Vector Machines (SVMs) are widely used for classification but are sensitive to mislabeled training data.
• Mislabeled examples can negatively impact SVM decision boundaries and performance on new data.

Purpose of the Study:

• To develop a novel, data-driven method for filtering atypical data instances prior to SVM training.
• To enhance the robustness of SVM classifiers against noisy datasets.

Main Methods:

• A new method based on L1-norm principal-component analysis and geometry is proposed.
• The method filters atypical data instances on a class-by-class basis in an unsupervised manner.
• The approach is computationally efficient and data-driven (touch-free).

Main Results:

• The L1-norm curation method effectively identifies and filters out atypical data instances.
• Experimental studies on real datasets demonstrate the method's efficacy.
• Protected SVM models show improved performance by mitigating the impact of data faults.

Conclusions:

• The proposed L1-norm based data curation method provides robust support-vector candidates for SVMs.
• This unsupervised, efficient technique enhances SVM performance by ensuring cleaner training data.
• The method offers a practical solution for improving classifier reliability in real-world applications.