Name		Name	Last commit message	Last commit date
Latest commit History 4 Commits
chizhu_nurbs		chizhu_nurbs
snake		snake
README.md		README.md
开源方案.pdf		开源方案.pdf

Repository files navigation

HUAWEI-DIGIX-AgeGroup

2019 HUAWEI DIGIX Nurbs Solutions

Notes:

RAdam,AdamW,Lookahead,CycLearn在/snake/11. lstm-atten中

CTR模型的选择，只需要更换调用的模型，如FibiNet,FGCNN,xDeepFM，调用的是 deepctr 库

Meta Train在/snake/7和8中，参考的是Home Credit 17名和Elo Senkin大佬的Solutions

原生的adam与Lookahead配合较好，效果比较为adam<AdamW<RAdam<RAdam + Lookahead<AdamW + Lookahead<adam + lookahead，效率是纯AdamW最快

Graph特征构建在/snake/3. Graph Feature中，有一些特征较慢，可以注释掉，我们选择的建图为二分图建图方法

Sites: chizhu : https://github.com/chizhu nurbs: https://github.com/suncostanx deepctr : https://github.com/shenweichen/DeepCTR/