Development of a Data Mining Study

“I love it when a plan comes together!” – John Smith 1928 – 1994

Before a plan can come together, you have to make one. Here are some key points to consider during development of a data mining study.

• Defining the aims

  • Start by choosing the subject and relevant aspects (customers, widgets or potatoes)

  • Give thought to essential criteria and the phenomenon you will be predicting

  • The aims must lead to actions and should be very precise

• Listing the existing data

  • Start kicking over the rocks

  • Look everywhere, IT systems (CRM/ERP), Spreadsheets, archived data (past marketing campaigns)

  • Make like a salmon and swim upstream, drill into aggregated data – find the source (enhanced interrogation maybe?)

• Collecting the data

  • Now you make like a seagull. “Mine… mine… mine!”

  • Start down your above list and get it someplace useful (aka, a database)

  • Remember, one record (row) for each statistical event (one record, one potato)

• Exploring and preparing the data (welcome back… keep reading, you’ll understand)

  • Cleanup on aisle three – check the origin, replace missing or incorrect data

  • You will also need to find aberrant values or outliers that are too far from the normally accepted values – the 600 pound potato… unless it’s an Idaho spud

  • Consider reducing the number of categories for categorical data

• Population segmentation

  • Depending on the aims of your study you may need to segment the data-set

  • Check out Cluster Analysis for more info

• Drawing up and validating the predictive models

  • Now here’s the meat and potatoes – using techniques such as;

  • Calculation of a score – logistic regression, linear discriminant analysis and decision trees

  • Building and comparing models – checking error rates in a confusion matrix (yeah, it’s a thing) or by superimposing their lift curves or receiver operating characteristic (ROC) curves

• Synthesizing Predictive Models of Different Segments

  • This step is only necessary if the data-set was segmented

  • Compare the scores of the different segments

• Rinse and repeat – go back to exploring and preparing the data, refine the methods and repeat until the results are medium rare… er, satisfactory

• Deploying the models

  • Implement the models on a computer system so that users can… well, use it

• Training the model users

  • Users should know the aim, principles of the tools, the how (without drinking from the techno-jargon fire-hose), limitations such as, “these are decision support tools, not tools for automatic decision-making.”

• Monitoring the models

  • Make a one-off analysis of the results

  • Some tools – like the algorithms used to calculate credit scores – should be monitored continuously for both its correct operation and its use

• Enriching the models

  • Continuous improvement – sharpen the saw