Numerous applications require the ability to discriminate one class of signals, signatures or objects from another based upon a collection of measurable features. State-of-the-art methodologies that perform this type of classification include Support Vector Machines, Neural Networks, and Random Forest. The DRF greatly enhances classification capabilities and supplants the current state-of-the-art.
The Discriminant Random Forest combines advantages of several methodologies and techniques to produce lower classification error rates.
The DRF produces smaller trees and forests at its peak performance which consumes less memory, demonstrates more efficient training, increased analytical capability and increased throughput.
The DRF produces stronger forests which improves performance in detection of weak signals or signatures.
It utilizes multiple feature dimensions yielding a more robust and efficient classifier.
The nonparametric methodology minimizes the violation of underlying model assumptions
The DRF, like its predecessors, is suited to applications requiring discrimination between two classes of interest, such as medical imaging analyses, detection of radiological sources, hidden signal detection, marketing analyses, and intrusion detection for cyber-security. Because DRF produces significantly lower error rates, it may be particularly valuable for applications in which errors can prove costly, such as medical and financial.
A DRF toolbox has been developed and utilized for several real, critical detection applications, such as hidden signal detection. In each case, the DRF achieved superior performance to other approaches, including Support Vector Machines, Cost-Sensitive SVMs, and the conventional Random Forest. Enhancements of the technology are in progress and will be released as subsequent versions. The DRF toolbox is presently command-line driven, but can easily be adapted to a Graphic User Interface (GUI).