Improved Random Forest for Classification.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Random forest classifiers are widely used but can be computationally intensive.
• Feature selection and tree pruning are crucial for optimizing classifier performance.
• Accurate classification is vital in fields like medical imaging and materials science.

Purpose of the Study:

• To develop an improved random forest classifier that minimizes the number of trees required for accurate classification.
• To introduce a novel method for feature reduction and establish a theoretical upper limit for tree addition.
• To demonstrate the classifier's effectiveness on diverse benchmark, medical, and industrial datasets.

Main Methods:

• Iterative removal of unimportant features from the random forest.
• Formulation of a theoretical upper limit for adding trees to guarantee accuracy improvement.
• Convergence of the algorithm on a reduced, yet informative, feature set.

Main Results:

• The proposed algorithm achieves classification with a significantly reduced number of trees.
• Demonstrated efficacy on benchmark datasets, histopathological breast tissue images (mitotic nuclei detection), and dual-phase steel microstructures.
• Significant reduction in average classification error compared to existing methods.

Conclusions:

• The developed random forest classifier offers improved efficiency and accuracy.
• The method proves effective across various applications, including medical and industrial data analysis.
• Further addition of trees or reduction of features does not enhance classification performance beyond the proposed algorithm's convergence point.