Random forest construction with robust semisupervised node splitting.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Random Forests (RFs) are powerful classifiers but require substantial labeled training data.
• Performance limitations arise from node splitting procedures when data is scarce.

Purpose of the Study:

• To develop a semisupervised splitting method for constructing RFs with limited labeled data.
• To enhance node splitting robustness by incorporating unlabeled data.

Main Methods:

• Kernel-based density estimation for accurate node splitting quality measurement.
• A multiclass asymptotic mean integrated squared error criterion for optimal kernel bandwidth selection.
• Projection onto a low-dimensional subspace to mitigate the curse of dimensionality.
• A unified optimization framework for selecting subspace and separating hyperplane.

Main Results:

• The proposed semisupervised splitting method effectively overcomes data scarcity limitations in RFs.
• The algorithm demonstrates improved performance over conventional methods by avoiding overfitting.
• Demonstrated effectiveness in computer vision tasks like object categorization, face recognition, and image segmentation.

Conclusions:

• Semisupervised splitting offers a robust solution for Random Forest construction with limited labeled data.
• The method provides a significant advancement for machine learning applications facing data constraints.
• The approach shows broad applicability in computer vision and other domains requiring robust classification.