请直接查看guide.pdf了解赛题，该项目是初赛第一名的模型。
- scikit-learn
- tqdm
- pandas
- numpy
- scipy
- tensorFlow>=1.12.0 (最好不要超过1.14)
- Linux Ubuntu 16.04, 128G内存(64G应该足够)，一张显卡
可以直接运行
pip install -r requirements.txtmkdir data
cd data
#Download data from https://pan.baidu.com/s/1ASQMms_u70psRgW_KEyT2Q
#Password: burw
unzip algo.qq.com_641013010_testa.zip imps_log.zip user.zip
cd ..
mkdir testdata
cd testdata
#Download data from https://microsoft-my.sharepoint.com/personal/xiuniu_microsoft_com/_layouts/15/onedrive.aspx?id=%2Fpersonal%2Fxiuniu%5Fmicrosoft%5Fcom%2FDocuments%2FAISchoolAdsProject%2FTestDataV2&originalPath=aHR0cHM6Ly9taWNyb3NvZnQtbXkuc2hhcmVwb2ludC5jb20vOmY6L3AveGl1bml1L0VpQ0U2cElFNXVoR28zY2h3cm5xMF9NQk5FakViRHVsaXBDWDJ0eGJFanMxZ3c%5FcnRpbWU9RFNKNVFxVDYyRWc python src/preprocess.pypython src/extract_feature.py
python src/added_feature.pypython src/convert_format.py1）缺失值NA用0填充
2）将Word2Vec和DeepWalk得到的embedding拼接起来，并且掩盖到5%的广告
3）将需要用key-values的稠密特征正则化到[0,1]之间
mkdir results
mkdir submission
mkdir model_tmp/model/
cd model_tmp
mkdir model
cd..
python train.py
python postprocess.py CINmkdir results
mkdir stacking
python LGB.py
python postprocess.py LGB所有的结果会在训练结束之后显示，同时也会以如下格式存储在results/CINScore.txt 和LGBScore.txt文本文件中：
Time
Username's score is score_number.