Multilayer Perceptron (M L P)
The most famous of all supervised Neural networks, and justifiably so : at the present time, and properly handled, the MLP is the best of all known tools for regression and classification.
Yet, it is hard for the newcomer to the field to draw a clear picture of what MLPs really are, and how they really work, from sales litterature. This is rather unfortunate, as popular litterature already clouded the picture with hazy hype about "brain-like" techniques, and frequent references to Artificial Intelligence (which MLPs have nothing to do with).
The present hegemony of MLP in the Neural networks world is now threatened by the increasing popularity of another type of Neural networks, the RBF networks.
Multiple Linear Regression (M L R)
Please see here.
Neural networks (NN)
Certainly the most innovative set of techniques in Data Mining. This generic term covers many different techniques, but only a few of them have made it to the market so far.
In this short glossary, we cannot get to what NN actually are or how they work, but we'll touch a word on what they do (and, to be honest, what they don't do) . Well, they do just about anything in terms of predictive modeling, and they have the potential of doing just about anything in descriptive modeling as well.
Te be more specific :
1) In predictive modeling, most Neural networks (MLP, RBF...) do non parametric regression. As a consequence, they also do probabilistic classification, as the two problems are closely linked. As a matter of fact, MLP is just about the best regressor known at this time, and RBF is a close runner-up. Their main property is to be able to get close to a regression function that is not just a straight line, but an arbitrarily twisted curve.
Neural networks doing predictive modeling are usually refered to as "supervised networks".
2) In descriptive modeling,
only one type of Neural network is currently available on a commercial basis
: Kohonen maps,
used for both clustering
reduction (but more powerful Neural networks exist in the labs, that
outperform Kohonen maps in several respects).
Neural netwoks doing descriptive modeling are usually refered to as "unsupervised networks".
Now for what Neural networks cannot do. NN incorporate parameters
that need to be adjusted through a training
phase, but unfortunately, the values of these parameters carry no
useful information about the process that generated the data. Therefore, even
a particularly well trained NN can do very well what it is supposed
to do, but he won't give you additional insight about your data. NN are
said to be "black boxes". In other words, there is no interpretation
of a NN.
In addition, building a good NN takes both time, practice and a reasonably good understanding of its inner workings.
Neural networks are now commonplace in Data Mining software, but they often are underused, or worse, misused. Although NN are in no way miracle tools, they are irreplaceable when high accuracy results are mandatory.
An other common name for categorical variable.
Please see here.
A variable is numerical (or continuous, or "interval") if its value can be any real number.
Typical examples of numerical variables are "Age", "Weight", "Revenue", "Speed".
See also : Categorical, Ordinal, Binary.