Discriminative orthogonal neighborhood-preserving projections for classification

Area of Science:

• Machine Learning
• Data Analysis
• Dimensionality Reduction

Background:

• Orthogonal Neighborhood-Preserving Projection (ONPP) is an effective orthogonal linear algorithm for high-dimensional data analysis.
• ONPP addresses the out-of-sample problem inherent in manifold learning algorithms like Locally Linear Embedding.
• However, ONPP's supervised classification performance is limited as it overlooks inter-class relationships.

Purpose of the Study:

• To introduce Discriminative ONPP (DONPP), an enhanced algorithm for supervised classification.
• To develop a semisupervised version, Semisupervised DONPP (SDONPP), leveraging unlabeled data.
• To improve classification accuracy by incorporating both intra- and inter-class geometric information.

Main Methods:

• DONPP integrates intraclass and interclass geometries, considering neighborhood information between different classes.
• DONPP maintains the orthogonality property of the original ONPP algorithm.
• SDONPP extends DONPP by utilizing unlabeled samples to boost classification performance.

Main Results:

• DONPP effectively utilizes both intraclass and interclass geometric information for improved classification.
• SDONPP demonstrates enhanced classification accuracy by incorporating semisupervised learning principles.
• Empirical studies validate the superior performance of both DONPP and SDONPP over existing methods.

Conclusions:

• DONPP offers a significant advancement in supervised classification by considering comprehensive data geometry.
• SDONPP provides a robust semisupervised approach, further boosting classification accuracy.
• Both DONPP and SDONPP represent effective extensions of ONPP for classification tasks.