Software

Salford Systems is releasing TreeNet(tm) Stochastic Gradient Boosting in collaboration with Stanford University Professor Jerome H. Friedman. TreeNet, designed for both classification and regression problems, is Dr. Friedman's latest contribution to data mining technology. TreeNet has shown remarkable promise in text mining as well as numeric data mining. The technology combines the use of small CART(R) decision trees, stagewise use of small training data subsamples, adaptive rejection sampling, very slow learning rates, and redefinition of the target variable to achieve new levels of generalization accuracy. TreeNet is robust in the presence of common data errors including missing values, outliers in predictors or the target, and mislabeling of the target. Independent tests show TreeNet to be unusually resistant to overtraining.

The software offers a collection of new visualization tools for this class of technology, including 2-D and 3-D derivative plots, interaction diagnostics, class specific as well as overall variable importance scores, and standard performance summaries such as gain charts, ROC curves, and prediction success tables.

TreeNet is the first data mining tool to be embedded in Salford Systems' new data mining platform. CART, MARS and other new data mining, data management, and data exploration tools will be incorporated into the platform later this year.

Further detail on TreeNet stochastic gradient boosting and the software can be found at http://www.salford-systems.com/treenet.html. The software is available for immediate evaluation and academics may request teaching and research versions.