Data Science: The Hard Parts

Chapter 6. What’s in a Lift?

There are very simple techniques that help you accomplish many different tasks. Lifts are one of those tools. Unfortunately, many data scientists don’t understand lifts or haven’t seen their usefulness. This short chapter will help you master them.

Lifts Defined

Generally speaking, a lift is the ratio of an aggregate metric for one group to another. The most common aggregation method is taking averages, as these are the natural sample estimates for expected values. You’ll see some examples in this chapter.

$$\text{Lift}(\text{metric},A,B)=\frac{\text{Metric}\text{aggregate}\text{for}\text{group}A}{\text{Metric}\text{aggregate}\text{for}\text{group}B}$$

In the more classical data mining literature, the aggregate is a frequency or probability, and group A is a subset of group B, which is usually the population under study. The objective here is to measure the performance of a selection algorithm (for example, clustering or a classifier) relative to the population average.

Consider the lift of having women as CEOs in the US. Under a random selection baseline, there should be roughly 50% female CEOs. One study estimates this number at 32%. The lift of the current job market selection mechanism is 0.32/0.5 = 0.64, so women are underrepresented relative to the baseline population frequency.

As the name suggests, the lift measures how much the aggregate in one group increases or decreases relative to the baseline. A ratio larger or smaller than one is known as uplift or downlift, respectively. If there’s no lift, the ...