Ali Mobile Recommendation Algorithm Competition - http://tianchi.aliyun.com/competition/introduction.htm

0. Download the data as described in ./data_1/README.md and/or ./data_2/README.md.

1. Run TianChi3/main_preprocess.py once to pre-process the data

2. Run TianChi3/main_single_model.py step by step to tune and train the models.

The best single GBDT model can be obtained from TianChi3/main_single_model.py has F1 score around 10.4%, which can rank about 100 in Season 1.

To improve the performance:

1. By changing the time parameters in utils.gen_feats, and utils.gen_ic_ind_feats, more labeled data can be generated, which is useful because the dataset is highly unbalanced.

2. Add more time intervals in utils.gen_feats, and utils.gen_ic_ind_feats.

3. Cross-validation can be used to select better hyper-parameters.

4. Model ensemble can be used. A possilbe method:

   1. Train multiple 'not bad' single models with different algorithms (logistic regression, GBDT, adaboost, randomforest...).

   2. Use single models' outputs as input to train a new logistic regression.

   3. Use the outer logistic regression to make the prediction.