Chapter 6. Applying Data-Mining Predictions

In Chapter 5 we examined a data-mining model used by a fictional retailer named Savings Mart. This small retailer used the model to determine what factors most affected shipments to its stores. The model was designed to predict the number of days between shipments and the quantity for each store and vendor. The model was based on a decision tree algorithm. Processing the algorithm determined that Store ID, Vendor name, and Product type were the main factors determining what the number of days should be. Quantity was affected by the days since last shipped, Vendor name, and Store ID.

Now that the managers of Savings Mart are aware of the prediction results, they would like to redesign the way product orders and shipments are automatically generated. Their primary goal is to increase profits by reducing operational overhead. Previously, orders were generated based on minimum and maximum threshold quantities for each store and product. As purchases were made, the system would check to see whether the available quantity fell beneath the minimum threshold amount for that product. When this occurred, the system would automatically generate an order for that store with a quantity based on the maximum threshold amount.

The method for generating orders seemed like a good one when Savings Mart first began operating in 2001. Unfortunately, the company found that it resulted in the generation of shipments almost on a daily basis. The cost associated with shipments was high and thus reduced profitability for each store.

This chapter will examine code designed to generate new predictions and apply the results to a newly designed shipment methodology. The code is encased in a modified version of the LoadSampleData standalone Windows program introduced in Chapter 5. The modified version will utilize a new table containing the results of the processed mining model.

The chapter also includes a case study about a data-mining solution used to predict Web usage. The solution was utilized by a Web site portal that organized content for its subscribers. The solution examined requested pages and used the prediction results to make suggestions about other pages the user might be interested in.