Lift Charts and Why they are important in business
Driving factor for a business to run successfully is to have profits. If there are no profits, then business will eventually end.
We would see many advertisements saying “Since 989” etc., such companies are profitable because they target right audiences and hence they are able to sustain for longer time. During the start of every business, there will be lot of spending done in advertisements and sending offers to everyone and so on. But this cannot be continued for a longer duration because, if we spend 100$ overall for all audiences, everyone may not be interested in this offer or ad and as a result our returns will be very less compared to 100$ and hence we will be in loss.
Hence, if we have a business problem like, if I own a car company, who will buy my car and how many should I actually bring to market?. These questions can easily be answered by a Lift curve or Lift charts, If I have a model that scores the population on the buying rate.
Lift Curve tells us about how much benefit we get by targeting creamy layer than the whole population.
So coming on to the logic behind this, the inputs we need for lift chart is a scored population with actual value and predicted score.
We rank order the population based on predicted score and split the population into deciles or ventiles and get the cumulative mean of actuals over mean of actuals.
Lets assume, the below table is a scored population with actual class and predicted score
We need to sort the scored population in descending order and get the cumulative actuals like below:
Now the lift curve is defined in such a way that x-axis is the number of expected returns if I don’t have a model but simply responded at random.
Lift is calculated as If I choose 10 cases and use our model to predict 1’s then out of 10, we would be right 3 times. If we simply choose 10 cases at random them we expect 10*(7/20)=3.5 times. Then the gives us a lift of 3/3.5 ~ 0.85
A good classifier will give me a good lift with very less cases.
The above lift curve tells us that if I had to cover all 7 cases then I need to target full population, if I want to cover 4 cases, then we need to target 11 cases.
