Once you have successfully processed your mining structure and mining models, it is time to view your results. To begin viewing those results, simply click the Mining Model Viewer tab.
Within the Mining Model Viewer tab, select the mining model you wish to review. Ideally you gave your mining models meaningful names so you can easily distinguish mining models from the names in the pull down list. The actual screen shown depends on the algorithm used by your selected mining model.
For example, the screen print below shows the results of a Decision Tree mining model.
As the name suggests, Decision Tree results show a tree with branches and leaves. Darker shaded branches contain more cases than lighter shaded branches. Clicking through the tree enables a user to view the column value splits where branches formed. The algorithm used the cases and columns included in the mining model to determine significant branches and leaves. SSAS does not allow you to directly set any branch values which both limits user control and avoids user bias.
Although not included in the screen print, the Mining Model Viewer includes a legend showing the number of cases found in each branch and the probabilities of each value found in the branch.
While traversing the tree provides detailed insights into the data, a more summarized view of the results may deliver a sufficient understanding. Using the same mining model, click on the Dependency Network tab then click on your predicted value. The Mining Model Viewer in the following screen print shows the input columns correlated to the predicted value.
To easily determine the strongest correlations, slowly move the slider on the left of the screen from All Links downward to Strongest Links. The last input column connected to the predicted value is the strongest correlation. Just remember that even a strong correlation does not necessarily mean the input column causes the actual outcome of the predicted value.
The Mining Model Viewer shows different results and options for each algorithm. For example, had I chosen a Clustering or Naïve Bayes mining model instead of the Decision Tree I selected, the screen prints above would look very different. In fact, the predicted value for each case could vary from algorithm to algorithm even if you include all the same cases and input variables.
If the results can vary from mining model to mining model, how do you determine which mining model produces the most accurate predictions? SSAS helps you decide in graphical fashion.
Once you have reviewed your different mining model results, you are ready to compare those results to each other. Click on the Mining Accuracy Chart tab. Select the Predict Value that interests you. Notice the Use mining model test cases button. That button allows SSAS to test the predictions of your mining models against the cases set aside when you first created the mining structure. You can test your mining models against a different data set, but by randomly holding out cases from your initial population, SSAS simplifies your comparisons.
Click the Lift Chart tab. The screen print below shows a number of colorful lines on a chart of Target Population % versus Overall Population %. Also shown are scores for each line in the legend overlaying a portion of the chart.
The blue 45-degrees line represents a random guess. My predicted value could be either Yes or No. To find 100% of my targeted Yes outcomes (aka Target Population), I would have to search all 100% possible outcomes (aka Overall Population).
The pink line climbing at a sharp angle represents perfection. Because approximately 5% of my Overall Population had an actual Yes outcome, I would need to search only 5% of the Overall Population if I already knew the outcomes.
Ideally, your mining models will appear closer to the perfect line than the random line. The green line closest to the perfect line is my Decision Tree model. As the lift chart shows, using my Decision Tree mining model, I can find more than 80% of my Target Population while searching only 20% of my Overall Population. Searching 20% of the Overall Population, my Clustering mining model found almost 70% of my Target Population and my Naïve Bayes mining model found almost 50% of my Target Population. Even the simplified Naïve Bayes mining model which cannot use continuous values still delivers far better results than the 20% random selection.
Finding more of your Target Population in less of your Overall Population means savings. Using a trusted prediction to take an action that alters an outcome is the real power of data I mentioned in Part I.
How to get details for the chosen 20% of the Overall Population for action? I will walk through a simple export example in the final part of this series.