AVAZU-CTR-Prediction

In online advertising, click-through rate (CTR) is a very important metric for evaluating ad performance. As a result, click prediction systems are essential and widely used for sponsored search and real-time bidding.

For this competition, we have provided 11 days worth of Avazu data to build and test prediction models. Can you find a strategy that beats standard classification algorithms? The winning models from this competition will be released under an open-source license.

Final solution:

1. folder: Summary / MAIN_IMPROVE
2. results: 0.3882
3. ranking: 55/1786

Tips:

1. hour: 14-10-21-00 ~ 14-10-30-23 => hour, date => night, day, holiday, weekday
2. unpreprocessed benchmark random forest
3. feature engineering
   • one-hotted (num, less than 10k distinct value)
   • hash (cate)
   • feature interaction
   • categorical less than 10 times transformation
   • long tail feature log-transformation
4. python fast learning (gradient decent)
5. xgboost (gradient boosting)
6. vw online learning (log-loss)
7. L1, L2 tuning
8. meta data
9. ensemble models

latest findings

31.10 = Halloween being qualitatively different from the rest of the sample

maybe training specifically on 24 (Friday) and/or 25-26 (weekend ~ holiday) would help? Holdout 2 days of data for CV, just 1 gives poorer performance (around 0.006 as you mentioned above) Split the data in one file per day Refactor the training and validation logic so I can pass a range of days I want to train/validate on.

http://cran.at.r-project.org/web/packages/FeatureHashing/index.html

preprocessing

1. add columns for Hour , Weekday , Public Holiday
2. Hash / one-hotted
3. Split train based on different days

modeling

1. python fast learning
2. vw
3. xgboost
4. caret adaptive (svm)
5. solution below
6. ensembles
7. calibration (-0.000x)

solution

1. GBDT - feature generation
2. Hashing tricks - feature generation for FM(Factorization Machine)
3. Factorization Machine