This adjustment process continues until the network produces output sufficiently close to the actual output. They can take two forms:
- Similar to regression analysis or discriminant analysis. The input dataset consists of records with known outcomes (i.e. the targeted respondent purchases product or does not purchase product). The neural net produces predicted output values that it compares with the known outcomes. The trained model can be saved and used to predict purchase behavior of future respondents.
- Similar to cluster analysis or factor analysis. The input dataset does not have known outcomes. The neural net will group input patterns together and is useful in identifying similar segments in a market.
Logistic regression is a technique that identifies variables that are important for distinguishing between two groups of individuals. It should be mentioned that other techniques such as regression and discriminant analysis can also be used to predict a dichotomous dependent variable.
However, these techniques depend on certain assumptions (i.e., multivariate normality of the independent variables and equal variancecovariance matrices for the two groups). When these assumptions are violated (which they typically are in studies of this type), the results of the analysis may be less than accurate. Logistic regression does not depend on these assumptions.This method implicitly minimizes the diversity within each segment - and thus, in this case, produces distinct segments with homogenous needs, preferences, etc.
Factor analysis is a statistical technique that is used to identify the structure within a set of variables. By examining the association among variables, factor analytic techniques produce a smaller set of variables or factors that represent the underlying dimensions of the original set of variables.
Each factor is not a single, directly measurable entity, but rather a construct that is derived from the measurement of the original set of variables. This technique is often used for the purpose of data reduction- that is, reducing a large number of variables to a smaller set of factors greatly simplifies the description and understanding of large sets of data.
Neural networks are an alternative to traditional statistical techniques for prediction, classification, segmentation, and time series analysis. A primary advantage of neural networks is that they can find non-linear relationships in the data.
They do not depend upon the same assumptions (i.e. multivariate normal distributions, equal variance-covariance matrices, etc.) as conventional techniques.
Neural networks "learn" patterns in the data, and use an iterative process to create models. They start with randomly generated weights and compare these to known predicted outputs. The weights are then adjusted and compared again. This adjustment process continues until the network produces output sufficiently close to the actual output. They can take two forms: