Enriched Random Forest for High Dimensional Genomic Data.

Area of Science:

• Machine Learning
• Bioinformatics
• Computational Biology

Background:

• Random forest algorithms excel with high-dimensional data but struggle when features vastly outnumber samples and informative features are rare.
• Traditional random forest performance degrades significantly under conditions of extreme feature-to-sample ratios and low proportions of relevant features.

Purpose of the Study:

• To develop a novel ensemble method, Enriched Random Forest, to enhance the performance of traditional random forest on high-dimensional datasets.
• To improve prediction accuracy by reducing the influence of trees utilizing less informative features.

Main Methods:

• Proposed a modified random forest algorithm, Enriched Random Forest, employing weighted random sampling for subset selection at each node.
• Evaluated the approach on high-dimensional microarray datasets for both regression and classification tasks.
• Demonstrated the utility of balanced leave-one-out cross-validation for reducing computational load and sample size during feature weight computation.

Main Results:

• Enriched Random Forest demonstrated improved prediction accuracy compared to traditional random forest.
• The enhancement was particularly notable in scenarios with a very small percentage of truly informative features.
• Balanced leave-one-out cross-validation proved effective for efficient feature weight calculation.

Conclusions:

• Enriched Random Forest offers a significant improvement over traditional random forest for high-dimensional data analysis.
• The method is especially beneficial for datasets characterized by a large number of features and a limited number of informative ones.
• The proposed approach provides a robust solution for enhancing predictive modeling in bioinformatics and other high-dimensional domains.