Skip to main content
Established 2009

Artificial Neural Network Modeling

Over the past decade or so, machine learning techniques have allowed modelers to transcend traditional statistical techniques. Machine learning allows the modeler to test multiple potentially-relevant variables at once to test for relevance and proper variable weighting.
For example, supervised machine learning techniques allows the model to train itself by running itself for thousands of cycles with different combinations of variables, looking for the precise combination that reduces the error difference between predicted and actual values for the historical data set.

Meanwhile, unsupervised machine learning techniques can search for patterns and then cluster records into categories are groups in waysthat could never be determined by manual or even statistical analysis alone.

Artificial Neural Network Modeling

One of the machine learning methods that we at MindEcology apply to analyzing customer data is that of the artificial neural network. This is a computational technique that was inspired by the way that the human brain is constructed. The typical process involves first “training” the network and then, once trained, applying the new model to a set of as-yet-unprocessed data.

Training a neural network involves combining various input variables, modifying each one using “hidden” weight, and then creating an output. In what is called “supervised” learning, the output is compared to an actual historical variable and the difference (or error value) for each record is measured. The network process is then repeated thousands of times, with the network algorithm self-adjusting the hidden weight in order to try to minimize the difference between the actual (historical) and predicted value. The resulting network configuration that produces the lowest error value is selected as the final model.

In the context of building best customer profiles or site selection, training the artificial neural network involves processing various attributes of historical customers (or, in the case of site selection, of the demographic attributes of the households within the geographies of current locations). Then, once the network has been properly trained, prospect data (or, for site selection, proposed location-related data) is fed through the trained network. The result is a scored and ranked list of each of the prospects, with the score indicating which ones are the most attractive for the business or marketing challenge at hand.

Contact MindEcology today to get started.