Adaptive data embedding framework for multiclass classification.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Supervised manifold embedding is crucial for effective data representation in classification tasks.
• Existing methods may struggle with complex, nonlinear data distributions and noisy datasets.
• A need exists for adaptive and automated model generation frameworks.

Purpose of the Study:

• To design an engine for automatic generation of supervised manifold embedding models.
• To propose a modular and adaptive data embedding framework for classification (DEFC).
• To introduce novel concepts for controlling local data sample positions.

Main Methods:

• The DEFC framework involves data preprocessing, relation feature generation, and embedding computation.
• Introduced concepts of "friend closeness" and "enemy dispersion" for local data sample control.
• Employed bilevel evolutionary optimization for model identification and parameter tuning.

Main Results:

• DEFC effectively controls intraclass and interclass data sample positions.
• Demonstrated effectiveness on noisy synthetic datasets with nonlinear distributions.
• Validated performance using diverse real-world datasets from various application fields.

Conclusions:

• The DEFC framework offers an effective approach to automatic supervised manifold embedding model generation.
• The introduced concepts generalize and extend the Fisher criterion for improved data embedding.
• DEFC shows promise for enhancing classification performance across different data types.