Editor’s note: David A. Bryant is vice president at Ironwood Insights Group.
Earlier this year, I wrote a two-part article covering machine learning techniques – cluster analysis and decision tree analysis. This article will cover a third machine learning technique that is commonly used by market researchers: artificial neural networks (ANNs).
The two types of ANNs that are frequently used in market research include multilayer perceptron and the radial basis function. Each has their differences and advantages.
In this article I am going to focus on the multilayer perceptron function because of its ability to solve linear and nonlinear problems. I will avoid going into too much detail regarding the mathematics involved in calculating the neuron weights in a multilayer perceptron (MLP) model.
Two examples of scenarios using the multilayer perceptron procedure include:
Let’s explore the multilayer perceptron algorithm using telephone customer churn data.1
ANNs use an unsupervised machine-learning algorithm to identify patterns found within unlabeled input data. The MLP procedure produces a predictive model for one or more dependent variables based on the values of the predictor variables. The MLP procedure is robust because the dependent variables and the predictor variables can be categorical or continuous or any combination of both.
Figure 1 shows that every multilevel perceptron consists of three types of layers – the input layer, the output layer and the hidden layer. The hidden layer can consist of one or more layers of neurons.
The input layer receives the initial data to be processed. The required task, such as a forecast of who will leave the telephone company in the near future, is performed by the output layer. The hidden layers of neurons are the true computational engine of the MLP.
MLPs are composed of neurons called perceptrons. So, before g...