http://www.data-mining-cup.com/2003/Wettbewerb/Aufgabe/1190662447/

Problem:

• Minimize spam email through filter: 8000 training, 11000 test records
• Contest places constraint that no more than 1% of real mail is thrown away

Methodology:

1. 1. Import data straight from contest
2. 2. Optimize parameters using only training records
3. 3. Build 400 random trees on 8000-sample training data, ap=0.05
4. 4. Allow emails with 50% confidence or higher that email is non-spam
5. 5. Complete analysis took less than 1 hour

Results:

• Our Spam Thru: 53 (1st place had 31, 2nd place had 68)
• No Spam Out(%) 0.66, i.e. <1% of real mail is thrown away
• We would have ranked 2nd out of 514 participants and 86 entries

Data Mining Cup 2003 Results