Home → Techniques and Tips → NeuralTools → Correlated Variables and Interactions between Variables
Applies to: NeuralTools, all releases
Does the software create an optimally "simple" model by combining correlated variables (dimension reduction)?
Correlations between variables are much less of an issue with neural networks than they are with linear regression. Therefore, at this point NeuralTools does not address them. "Multiple-layer feedforward networks generally handle even massive amounts of correlation without complaints. Probabilistic neural networks can be hindered to some degree, as groups of redundant variables exert undue influence on the decision process" (T. Masters, Advanced Algorithms for Neural Networks, 1995, p. 294).
The first step would be to determine if correlations are present, say by using StatTools Correlation and Covariance analysis. If there is a significant amount of correlation in the data, that would point towards the use of MLF nets. However, a PN net could also be a viable option, if it makes accurate predictions on the testing data despite the presence of correlations.
Interactions between variables are often present within data. This situation exists when the relationships between the values of one variable vary when measured against the values of another variable. For example, the driving experience of males vs. females does not remain constant across all ages.
How does NeuralTools adjust the impact of any given variable for any interactions that are present? Does the software provide graphs (sometimes referred to as panel graphs) that would help the user visualize and explain any interactions?
Variable interactions present a problem with linear regression, but do not present a problem with neural networks. Therefore NeuralTools does not provide tools for analyzing variable interactions.
Last edited: 2015-09-03