The lift curve
1. The lift curve
In addition to the cumulative gains curve, the lift curve is a widely used visualisation of model performance.2. Lift curve construction
Constructing a lift curve follows a similar process as constructing a cumulative gains curve. First, you order all the observations according to the model. On the horizontal axis of the lift curve, you indicate which percentage of the observations is considered. On the vertical axis, the lift curve indicates how many times more than average targets are included in this group. Consider for instance the lift at 10%, and assume that the top 10% observations contains 20% targets. If the average percentage of targets is 5%, the lift is 4, because 20% is 4 times 5%. As another example, consider the lift at 50%, and assume that the top 50% observations contains 10% targets. As 10% is 2 times 5%, the average percentage of targets, the lift is 2 at 50%. A random model has a more or less equal percentage of targets for each group, and therefore the baseline is 1.3. Lift curve interpretation
Better models have higher lifts. Therefore, curves that are higher, have better accuracy. However, as for cumulative gains curves, also lift curves of different models can cross each other. Consider the example given here: model 2 is higher at 10%, but model 1 is higher at 80%. In that case it is hard to say which model is best, as it depends on the situation. If you can target 10% of the population, model 2 is better suited because you can reach more targets, whereas model 1 is better if you want to target a larger part of the population.4. The lift curve in Python
Constructing lift curves in Python is similar to building cumulative gains curves. First you need to make sure that the scikitplot as well as the matplotlib pyplot module are imported. Then, you can use the plot_lift_curve with two arguments. First you add an array with the true values of the target, and the second argument has the model predictions for the population.5. The lift curve in Python
The plot is given here. As we want to predict targets, you should only take into account the curve for class one.6. Let's practice!
Let's practice!Create Your Free Account
or
By continuing, you accept our Terms of Use, our Privacy Policy and that your data is stored in the USA.