Typically, we follow four steps to perform data mining: creating a dataset, selecting variables, developing the model, and deploying the model:
Let us try doing some data mining and predictive analysis following these steps. Let's say we need to prepare a back-to-school marketing campaign by sending out promotional mail. Our resources are limited, so we want to reduce costs by sending only to those that are most likely to respond. We have customers' demographic information and their previous response records from the last campaign.
Suppose we have the following demographic information: Age, Education
Level, Gender, and Household
Count. In addition, we have the response records of the last campaign. In our data, the previous campaign shows 1,002 positive responses out of 5,612 customer orders:
Selecting variables needs both domain expertise (experience) and statistics knowledge. We use our experience to pick the most influential factors, and use statistics to take care of those variables that would cause model-estimating problems, for example, those variables that cause multicollinearity. In this process, we may need to do principal component analysis, or variable transformations. For our example, we choose all the variables.
In this step, we need to decide what kind of model to use: decision tree, regression model (linear, non-linear, which functional form), or neural network, and so on, then do model estimation.
Developing a model needs expertise. As a quantitative researcher who worked for many years in econometric modeling with business and financial data, I know this task can be very tricky and complicated. In many cases, business analysts find it more productive to outsource this task to responsible and well-trained statisticians or economists.
In this example, we assume that this model developing work is outsourced. An economist used a third-party application (R, SAS, Stata, or EViews, and so on), connected to the dataset and built a Neural Network model. He exported the model as a PMML file, and sent it back to us. We simply need to import this PMML file:
We may want to have a closer look at the imported model. MicroStrategy can give us details of this neural network model. Here is how to do it: right-click the Response Predictor (Imported) (Scoring) metric and select Edit. In the Metric Editor, on the Tools menu, select View Predictive Model. A graphical representation of the model appears:
After importing the model, we may want to verify how well this model predicts before deploying it. This task can be done by applying this model on a dataset with known target variable values. We run the model, and compare actual target variable values with the estimated values. This process is model validation. In our example, we can create this dataset with known target variable values by simply using a filter, to include the data between 8/1/2013 and 9/30/2013, in which period responses to previous promotion are known:
Create a new report, add the Customer attribute, and add the following metrics: Back-to-School Responder, Response Predictor (Imported) (Scoring), Response Predictor (Imported) (Confidence). Run the report. Now we can compare the actual response with the predicted response side by side:
After model verification, if we are satisfied with the predicting accuracy of the imported model, we can deploy it. That is, we can apply the model to the modeling dataset (the dataset with the target variable unknown). In our example, we can create such a modeling dataset by changing First Order Date (ID) to after 9/30/2014:
Run the report. Now we get a list of 23 customers we should send promotional mail to:
