Lift Curve

[Atabarezz]

Here is the lift chart output, the chart starts at (0.1,0.1) not at the origin of (0,0) is there a reason to that?

 

Here are some scaling on both X and Y axis...

If we put eh point to the origin and redraw the cumulative response, is this a more accurate graph?

 The area under the curve (AUC or ROC metric) does it include the drawn area as well or not?

Best

[josethomas429 · Accepted solution]

This question is old and I am not with Alteryx support, but I was doing something related.  I thought I would comment.

I believe the AUC reported does indeed include the area between 0 and 0.1 even though it is not shown.

I checked this using the gains chart that is also provide in the output.  So say Alteryx gives you a gains chart like this:

Decile

Cum_Pct

0	g0
10	g1
20	g2
30	g3
40	g4
50	g5
60	g6
70	g7
80	g8
90	g9

100

g10

g0-g10 are numbers between 0 and 1.

The x-coordinates of the points in the curve are Decile / 100.  The y-coordinates of points in the curve are g0 - g10 (though point (0,g0) is omitted as you pointed out.

Using the trapezoidal rule, the AUC including the area that's 'falling off the picture' is like this:

AUC ~= sum(i= 1 to i=10) {((Decile[i] - Decile[i-1])/100)*(g[i] + g[i-1])/2}

It may look a little funky, but this calculation is easy to do in Excel.  Doing that in my own example, I can match the AUC Alteryx is outputting to 3 decimal places.  So I think it's the case that they're including the 'falling of the chart' area.  Why they didn't start the chart at 0 for the picture, I don't know.  I speculate that the incremental option would logically start at 0.1 and they decided to start both versions of the chart at the same place.

Best,

Jose

[josethomas429]

Oh noticed one other thing.  You mentioned ROC in your question and this is not an ROC curve.  The AUC reported by the Lift tool is the AUC under the Gains curve.  I think there are other Alteryx tools that calculate AUC for the ROC curve but I haven't used them.