CHAID vs ANN
Artificial neural networks (ANN) are "computational models inspired by animals' central nervous systems that are capable of machine learning and pattern recognition". They are presented as systems of interconnected neurons that can compute values from inputs by feeding information through the network.
Commonly, a class of statistical models may be called "neural" if they
- consist of sets of adaptive weights, i.e. numerical parameters that are tuned by a learning algorithm, and
- are capable of approximating non-linear functions of their inputs.
The adaptive weights are conceptually connection strengths between neurons, which are activated during training and prediction.
As aforementioned, marketers should divide the market into segments. A product offer would then be made to those people who are considered to be in the market segment for which the product was meant. Automation segmentation techniques have been used to divide the customers into a number of segments, like CHi-squared Automatic Interaction Detection (CHAID). CHAID is a type of decision tree technique, that is used for prediction as well as classification and detection of interaction between marketing variables. In practice, CHAID is often used in the context of direct marketing to select groups of consumers and predict how their responses to some variables affect other variables.
When a segmentation method such as CHAID is used to determine target segments which are more likely to respond to marketing actions, the data set is split into disjoint groups and the groups that are considered to be most likely to respond are targeted through a marketing action, e.g. mailed preferentially. Neural networks, however, do not segment data into disjoint groups. That means, a different approach is needed to develop target selection models by using neural networks. Since target selection models try to differentiate responders from non-responders, marketers could consider formulating the problem as a classification problem and develop a network that classifies each input pattern as a responder or non-responder. However, the classes are not well separated in target selection problems.
Furthermore, the miss-classification costs are asymmetric. Miss-classifying a possible responder as a non-responder and not targeting that individual is more costly to the direct marketing company than miss-classifying a non-responder as a responder. Therefore, another method than pure classification should be used. One solution could be to define a measure for the "likelihood" of an individual to respond (call-to-action) via the marketing action, e.g. mailing. The individuals are then ordered according to this measure, and only the ones that score above a given threshold θj are targeted (target scoring).
Artificial neural network models for target selection fall into this category. The output of the neural network is then a measure of likelihood of response (Potharst et al., 2001). Now, the goal of determining the neural network would be to determine a correct set of network parameters (weights) such that a good indication of the likelihood of response of the supporters is obtained, given the inputs to the network. Naturally, these inputs must convey information about the characteristics of the supporters, including demographic, geographic, or even psychographic and behavioral variables.
ANN configuration and analysis
In target selection problems we form the model of the likelihood of response at a single point in time, given various supporter characteristics and/or their past behavior. In real life, feed-forward neural networks are sufficient for target selection problems. The nonlinear activation function of the neurons also has an influence on the final neural network model for a given network structure and the number of neurons. The logistic sigmoid function:
is often used since the neural network output is expected to lie in the range [0,1] (likelihood of response). Thus, the neural network model can be analyzed as a generalized nonlinear regression model with the likelihood [0,1] of response to our marketing action as an output .
We may visualize our target selection model through a hit probability chart.
|Potharst et al. 2001|
A hit probability chart exposes what percentage of a selected population may respond to our marketing action. The x-axis shows a percentage of the total group selected for targeting and the y-axis shows what percentage of that group is a responder to our call-to-action. The hit probability chart for a successful model starts with high values of hit probability. As the size of the selected group increases, the percentage of responders within the selected group will decrease until the hit percentage is equal to the percentage of responders within the total marketing action eg. mailing considered in the campaign.
In real life, after data preparation and data set selection, marketers may use IBM SPSS Neural Networks as a modern modelling tool.
With SPSS Neural Networks, marketers may select either the Multilayer Perceptron (MLP) or Radial Basis Function (RBF) procedure. Both of these are map relationship techniques implied by the data. Both use feedforward architectures, meaning that data moves in only one direction, from the input nodes through the hidden layer of nodes to the output nodes. Prior research has indicated that neural network models may perform better than the CHAID model, while they perform on an equal level with the logistic regression models we aforementioned.
In general, ANN for Marketing can be viewed as a modelling tool that helps marketers to take decisions. ANN can recognize patterns, pick up key information that are not easily identifiable and develop relationships among them. Marketers are expected to fully take advantage of the ANN modeling in order to solve their target selection problem faster and more efficiently.