Lift is computed by comparing performance with a random selection model. I'll explain with your example below,
- assume that we didn't have any statistical/ML model for ranking/scoring the respondents.
- In that case assume we did a random ordering of respondents.
- A decile (10% of total population) is expected to have 10% of the respondents. In your case, there should've been (approximately) 488 respondents in 2500 cases.
- But after ordering the cases by score, you are seeing 44.71% of the cases in first decile against expected 10% (in random/no model case). This gives the gain of 44.71/10 = 4.471.
- For next decile, cumulatively you have covered 20% of the cases. You'd expect a random/no model scenario covers 20% of the respondents. But using scores, we covered 80% of them. That gives a cumulative lift of 80/20 = 4.