Testing Accuracy with Lift Charts (Basic Data Mining Tutorial)
Applies To: SQL Server 2016 Preview
On the Mining Accuracy Chart tab of Data Mining Designer, you can calculate how well each of your models makes predictions, and compare the results of each model directly against the results of the other models. This method of comparison is referred to as a lift chart. Typically, the predictive accuracy of a mining model is measured by either lift or classification accuracy. For this tutorial we will use the lift chart only.
In this topic, you will perform the following tasks:
The first step in testing the accuracy of your mining models is to select the data source that you will use for testing. You will test how well the models perform against your testing data and then you will use them with external data.
Switch to the Mining Accuracy Chart tab in Data Mining Designer in SQL Server Data Tools (SSDT) and select the Input Selection tab.
In the Select data set to be used for Accuracy Chart group box, select Use mining structure test cases. This is the testing data that you set aside when you created the mining structure.
For more information on the other options, see Choose an Accuracy Chart Type and Set Chart Options.
To create an accuracy chart, you must define three things:
Which models should you include in the accuracy chart?
Which predictable attribute do you want to measure? Some models might have multiple targets, but each chart can measure only one outcome at a time.
To use a column as the Predictable Column Name in an accuracy chart, the columns must have the usage type of Predict or Predict Only. Also, the content type of the target column must be either Discrete or Discretized. In other words, you cannot measure accuracy against continuous numeric outputs using the lift chart.
Do you want to measure the model’s general accuracy, or its accuracy in predicting a particular value (such as [Bike Buyer] = ‘Yes’)
On the Input Selection tab of Data Mining Designer, under Select predictable mining model columns to show in the lift chart, select the checkbox for Synchronize Prediction Columns and Values.
In the Predictable Column Name column, verify that Bike Buyer is selected for each model.
In the Show column, select each of the models.
By default, all the models in the mining structure are selected. You can decide not to include a model, but for this tutorial leave all the models selected.
In the Predict Value column, select 1. The same value is automatically filled in for each model that has the same predictable column.
Select the Lift Chart tab.
When you click the tab, a prediction query is executed to get predictions for the test data, and the results are compared against the known values. The results are plotted on the graph.
If you specified a particular target outcome using the Predict Value option, the lift chart plots the results of random guesses and the results of an ideal model.
The random guess line shows how accurate the model would be without using any data to inform its predictions: that is, a 50-50 split between two outcomes. The lift chart helps you visualize how much better your model performs in comparison to a random guess.
The ideal model line represents the upper bound of accuracy. It shows you the maximum possible benefit you could achieve if your model always predicted accurately.
The mining models you created will usually fall between these two extremes. Any improvement from the random guess is considered to be lift.
Use the legend to locate the colored lines representing the Ideal Model and the Random Guess Model.
You'll notice that the TM_Decision_Tree model provides the greatest lift, outperforming both the Clustering and Naive Bayes models.
For an in-depth explanation of a lift chart similar to the one created in this lesson, see Lift Chart (Analysis Services - Data Mining).