Gain Chart

yogafire

how to make a "Gain Chart" using RM 5? is it the same with "lift chart" ?

but when I try to make a lift chart using my model it's not like the example of gain chart on my book.
does anyone have an idea?

thanks

steffen

Hi yogafire

I did not know what a gain chart is, google sent me to http://www.dtreg.com/LiftGainTable.htm.

@first-question:
I don't think this is possible.

@second-question:
No. Example: Direct-Marketing campaign with a binary target buy=yes/no. As far as I understand this charts ..
- gain chart plots the percentage of contacted customer  vs percentage of sales (buy=yes). The latter percentage is also called recall.
- lift charts plots the percentage of contacted customer vs the increase of sales in comparison to the baseline (which is in general the random model, percentage of customer buying across the whole set).It is calculated as (percentage of sales (buy=yes) for the contacted customers)/(percentage of sales for all customers).

You see, it is related, but not the same.

The Lift-Chart by Rapidminer does the following: It plots the confidences/scores of a model vs the sales percentage. For this the confidences are discretized into bins. This chart allows you to see whether your scores are well calibrated (i.e. whether the scores are a good approximation to the probability). If you mark "show cumulative labels", too, you additionally see the "percentage of sales"-plot.  Since the (binned) confidences are sorted on the x-axis, the result is again a gain chart (given that the above definition of gain and lift-chart are correct).

@third question:
I think that it is theoretically possible to do it in rapidminer by generating attributes representing x and y and then using the default plotting functions. However this would require coding (in java or groovy). I would write the result of my model (target + confidences) to a file and use R instead.

greetings,

steffen

PS: I find your way of asking questions ... irritating. Think again how "but when I try to make a lift chart using my model it's not like the example of gain chart on my book." sounds to another person. Second I'd always perform some research via google before asking. This is a friendly remark, no offense.

yogafire

thanks steffen for the explanation....
by the way I'd like to ask for apology for the way i asked a question. I'm sorry if you or anyone else felt offended. I'm just really desperate at that time.... :-[

so based on your explanation maybe I would have to make my gain chart manually using my java code... and in my country reference about ROC, gain chart and lift chart is very limited... so I have to find it out myself and my english is not so really good... because of that, maybe english is a burden for any indonesian student... like me of course for example.

thank you very much.

steffen

Hello yogafire

As I said: no offense . I apologize if it sounded too harsh, I tend to forget the language barrier. Sorry again !

Regarding your task:

• excellent source for roc-curves: http://home.comcast.net/~tom.fawcett/public_html/papers/ROC101.pdf
• source for lift- and gain-charts: http://www2.cs.uregina.ca/~hamilton/courses/831/notes/lift_chart/lift_chart.html

side question: did you find out which exact type of chart you want to create ?

If you are a beginner and think that data mining and statistics is "your thing", I strongly recommend to learn R. If it is not the algorithms, than for the plotting capabilities

good luck,

steffen

yogafire

OK, here's my story. and my story can answer your side question and this is the source of me being posted that question.

I'm a computer science student. I'm on final year of my study and I'm on the verge of finishing my minithesis about classification on educational data mining domain. I have evaluated all my model and give accuracy as a parameter for choosing model. but my professor advises me to go further and deeper in analysis and if possible, make a visual evaluation in my analysis (my professor is kind like... put it this way, if he says "if possible", it means a "must", unfortunately ). after weeks of searching, I found some educational data mining related papers (most of all the latest) use gain chart in their classification analysis, not just accuracy. then I begin searching again and again for gain chart... to be honest, I've been heard about ROC and lift chart before, but here in indonesia, there are (almost) no references about those. and that stuff is just like a completely new for me.

then I remember on RM there are facilities for generating ROC and lift chart, but my professor just insist on gain chart. so i came to this forum and finally post that rude question of desperate. because if I don't finish this by the end of this week, I would likely to postpone my graduation for about 4 months... it's a long time... that's why I'm really desperate and to be honest your words has shocked me a little bit, but I do understand... we're in forum, and no one knows what we really look like and what we really do. we only communicate bu text. after all I don't feel offended at all. by the way thanks a lot for your explanation.

and now my progress on integrating "Gain Chart" feature on my java code can be said as 70% done, and I'm just really tired because I haven't (really) slept for 3 days just to finish my minithesis. I hope you do understand, nice to meet you.

regards,

Dimas Yogatama

steffen

Thank you for sharing your story

All what's left for me to say:  All the best for your thesis and all the energy for you to finish on the time !

kind regards,

steffen