Replies to This Discussion

Permalink Reply by Steffen Springer on December 5, 2009 at 3:07pm

Hm ...

If I understand this chart correctly, then the lift is calucaled for each decile. Could it be that your ranking of scores is so good, that indeed most positive class items have been ranked to the top, so that less positive class items are available at the lower deciles ?

Have you looked at the roc-curve / auc-value for this ranking ?

Permalink Reply by Ralph Winters on December 6, 2009 at 2:50pm

The red line doesn't make any sense. The blue line is not the cumulative lift of the red line, Why are the lines going in different directions between 50 & 60% ?. However the blue line does make sense since the cumulative lift value should always be equal to 1 at 100%. Maybe the red line represents the gains chart at the individual deciles, but then the y-axis would be wrong. Could be an overlay. Which package is this from?

-Ralph Winters

Permalink Reply by Paul Wilson on December 6, 2009 at 6:43pm

Hi Ralph,

Yeah, I'm confused by the red line as well.
The blue one does make sense to me as well for the same reasons you mentioned.
I also thought the red one could be the lift score for each decile separately (i.e. non-cumulative lift), but I'm only used to looking at cumulative lines in lift charts so I'm not quite sure.

By the way what do you mean by "could be an overlay"?

The chart is done in KNIME.

http://knime.org/

Thanks for your help guys !!

Permalink Reply by Robert A. Nisbet on January 31, 2010 at 7:38pm

The graph is not explained properly in the legend. The blue line is the cumulative lift chart, which asymptotes at 1.00 at the 10th decile (100% of the cases). The red line appears to be the incremental or decile lift value. Sometimes this value is called the "Lift Index", or just the "Lift". The red line (decile lift) is greater than 1.0 only to about the 3rd decile, then is goes below 1.0 (the level of random expectation - via a coin toss). To the mass mail marketer (for example), this means that he should mail to only the top 3 deciles, or the top 30% of the scored list sorted in descending order. Beyond the 3rd decile, the model is not effective in predicting response at any level greater than the random expectation.

Permalink Reply by Paul Wilson on February 1, 2010 at 7:38am

Thanks for your reply Robert.

It kind of looks odd to see the red line (decile lift) below the random expectation (green line) for deciles 4 to 10.

Is this even possible?

I know cumulative lift always = random expectation at the end of decile 10 and never goes below random expectation. However I wasn't sure if that's true for (non-cumulative) decile lift.

Thanks