"The graphic above was created using predictions produced by the algorithm we implemented — a Decision Tree. We chose this algorithm because, despite how intimidating the graphic appears to be, we'll have you interpreting the results in just a few minutes!
The graphic illustrates which aspects of the factory are best at predicting if a team will be productive or not. When we say aspects, we mean variables like the date, department, team number, etc. In our case, we discovered that the "incentive" and the "smv" variables were the ones with the greatest influence on the final prediction. To provide context, the "incentive" variable represents the amount of financial incentive offered to motivate a particular course of action. The "smv" (Standard Minute Value) variable represents the time allocated for a specific task.
How do we know that these two variables are the most predictive? We know this because of the information included in the square boxes in the graphic! For example, if we focus on the top part of each box, we will see that those two variables are the ones that appear most frequently. In fact, there is only a single exception in the lower left of the graphic where the variable "number of workers" appears at the top.
You may be wondering why the algorithm is called a Decision Tree. Well, you can think of the graphic as an inverted tree with the top box, called the "root", representing the most predictive feature and the boxes at the bottom, called the "leaves", which provide the predictions. These final boxes (leaves) are the ones that tell us whether a team in a specific department was productive or not. The prediction is shown in the "class" component at the bottom of the box. Keep in mind, even though all boxes feature a "class" component, we are only interested in the classes found in a leaf, not before.
"But how do I know which path to take from the top box to the final ones?" you might ask. Good question! Basically, we need to use the thresholds associated with every variable at the top of every square box. We have to compare our inputs with those thresholds – starting from the root – until we reach the leaves at the bottom. When we say "input" we are referring to the corresponding "incentives" and "smv" values for a particular team. We can obtain these values from any date we choose in the dataset or we can select the values ourselves to test hypothetical situations. This is excellent for us since it means we can use the Decision Tree to predict the outcome for future and/or hypothetical scenarios so long as we have their respective incentive and SMV values.
We get it, that last part was a bit overwhelming, so to show you how all this works, let's use an example: a fictional date where the "incentive" is 22 and "smv" is 4.44 to predict if a team will be productive or not under these conditions. Starting from the root, we see that our first comparison will be 22 <= 22 when we substitute "incentive" for 22.
So, is 22 less than or equal to 22? Yes! Since the assertion is True, we follow the arrow to the lower left box. This is a universal rule of Decision Trees: if the assertion is True, we continue to the left; otherwise, if it's False, we continue to the right.
True = Left. False = Right.
Now we repeat the process for the rest of the boxes. We apply the same approach to the next box, which has the comparrison 4.44 <= 3.92 after we substitute "smv" for 4.44.
So, is 4.44 less than or equal to 3.92? No! Since the assertion is False, we follow the arrow to the lower right box. Again, after substituting "smv" for 4.44 in that box, our final comparison is 4.44 <= 4.865.
So, is 4.44 less than or equal to 4.865? Yes! Since the assertion is True, we follow the arrow to the final lower left box, the leaf, which predicts the team will be "Productive". In other words, if we have a team with an "incentive" of 22 and a SMV of 4.44, it will be a productive team!
To summarize, the Decision Tree not only told us which variables have the strongest predictive power ("incentive" and "smv"), but it also allows us to make predictions using values of our own choosing. This is the power of Decision Trees! Any questions?"
Using Random Forest¶
To confirm and validate the results from our Decision Tree model, let's use a Random Forest to compare results.
