Let's have a look at the many ways you can use DataRobot to get actionable insights, starting with the Profit Curve and Payoff Matrix (select the recommended model from the Leaderboard, and navigate to Evaluate > Profit Curve tab). As mentioned in Gut Check the Recommended Model, you can set the dollar amounts associated with a good or bad prediction. In this example, an accurately predicted on-time delivery saves $1, as does an accurately predicted late delivery. Incorrect predictions about whether the deliveries will be late or not cost $1.
Here, you can see with a probability threshold of 50%, the total profit associated with this model is $1,600. You can change the probability to .2, for example, and see how the profit changes. These displays give you and your stakeholders the opportunity to decide your risk tolerance for making a wrong prediction, and provide greater certainty about the costs associated with these predictions.
The meaning of “true negative” and the like can take some mind bending, so if you hover over the labels in the boxes, DataRobot will interpret the labels for you:
Let's have a look at a word cloud for the model (from the Models tab, select Insights > Word Cloud tab).
You see that the auto-populated model for the word cloud is an algorithm that processes text and is part of the blueprint for our "recommended model." If you click the dropdown, you can toggle between word clouds for the two text features (columns) in the dataset, Molecule/Test Type and Item Description.
In blue you see what’s more likely to be on time and in red you see what’s not. Take your time reviewing the word cloud to ensure you understand what it's telling you. For example, here we see the word determine, which turns out to be the brand name of an HIV test as well as the name of two HIV treatment drugs. You can hover over the word and see how many times it appears in the dataset.
So, as with the other insight features, the word cloud can be used to help isolate opportunities to reduce late deliveries. Perhaps, for example, you would consider switching to another distributor of the drugs whose deliveries are usually on time.
Now let's check out hotspots for this model (from Models tab, select Insights > Hotspots tab), which provide a simple list of possible rules that you can apply to reduce late shipments. (Note that the Model dropdown is populated automatically because it contains the algorithm that creates the “rules” identified in the Rule table below the Hotspot display.)
When you hover over the spot shown in the below image, you see that shipments by truck of line item quantities, or packs, of supplies greater than 1,689 are more likely to be late.
Perhaps for shipments of this size, you may consider a different delivery method.