Our award-winning data mining products are high performance supplementary tools that work well in conjunction with your neural networks. Users of artificial neural networks may even find that these tools are powerful and flexible enough to replace neural networks. An overview of how CART, MARS, and TreeNet® can be of value follows.
Using TreeNet® as an alternative to Neural Networks
TreeNet® is Salford System's version of a traditional neural network. We think of TreeNet® as a network of trees. For more detailed information, including a downloadable fully functional trial version of the software and a tutorial, see the TreeNet® product page.
What are the advantages of TreeNet® over a neural net?
TreeNet® is not sensitive to data errors and needs no time-consuming data preparation, preprocessing or imputation of missing values. TreeNet® is resistant to overtraining and is over 100 times faster than a neural net. Finally, TreeNet® is not troubled by hundreds or thousands of predictors.
Can a neural net do anything a TreeNet® cannot?
Yes. Version 1.0 of TreeNet® cannot accept more than one target variable at a time. To model a collection of targets a separate TreeNet® model must be developed for each target independently. Also, neural nets can simultaneously estimate a function and its derivatives whereas TreeNet® is not designed to estimate the target function derivatives.
Using the CART decision tree as a preprocessor
CART has been shown to be an effective variable selection and data preprocessor for neural networks. In the January/February 1998 issue of PCAI Magazine, Will Dwinnell describes how he used CART to reduce the number of inputs to obtain more accurate results in significantly shorter training times. He also experimented with using the CART decision tree outputs as inputs into his neural network to obtain even better results. The paper, Modeling Methodology 2: Model Input Selection, is available in PDF format.
A similar use of CART is reported in the August 2000 issue of PCAI Magazine, where Dr. Wayne Danter of Critical Outcome Technologies discusses his use of CART, MARS, and neural nets in drug discovery. The paper, Molecular Data Mining Tool: Advances in HIV Research, is available in PDF format. A number of other researchers have reported comparable results.
Benefits of using CART as a preprocessor for your neural network include:
- Very fast training times.
- No need to transform or prepare the data (CART can easily use raw data).
- Automatic handling of missing values.
- Automatic handling of categorical (nominal) predictors.
- Ability to handle very large numbers of predictors (up to 8,000).
- Ability to handle very large training data files.
Using MARS as an alternative to neural networks
MARS is an extraordinarily fast and flexible regression model builder that works by (1) breaking up every predictor into dynamically-discovered regions and fitting region specific regression slopes, and (2) discovering region specific interactions. A typical MARS model may involve the testing of between 100,000 and 10 million models, but the high speed shortcut algorithms yield very fast performance.
Serious comparisons of MARS and neural networks began in 1993 with the publication of Richard DeVeaux's article, A Comparison of Two Nonparametric Estimation Schemes: Mars and Neutral Networks in Computers in Chemical Engineering, Vol. 17, No. 8. DeVeaux and his coauthors found that the MARS models at were least as accurate as his neural networks and trained in a fraction of the time. Many subsequent researchers have confirmed these findings. See, for example, research conducted by Ajith Abraham, and now published in the Springer Verlag Lecture Notes in Computer Science: For more detailed information, including downloadable fully functional trial software and tutorials, see the MARS product page.