mind
2020 MIND news recomendation first place solution
I do not open source the code for the training loop using utils/melt gezi and husky so the code can not run, sorry(btw you can find some old code of melt from wenzheng/utils).
You can get most model part code from projects/ai/mind/src/model.py
training use sh ./train/v17/body.sh
tfrecord generated in projects/ai/mind/prepare
Code can run on both tf 1.14+ and tf 2.1+
Newst result with AUC 0.7187(Current STOA of MIND 20210104) explained:
我们在赛后刷新了MIND的最佳成绩(AUC 0.7187 20200104)并且在微信公众号介绍了所用的方法:
https://mp.weixin.qq.com/s/7_lSvNrCc_hJPb6giN7OJQ
Document of the MIND contest(AUC 0.7131):
https://docs.google.com/presentation/d/1TCcv4QoCv132ubd78xMxegEM7qRqIIRnF2Wj2oIFP94/edit?usp=sharing
BTW,目前榜一UniUM的成绩太强了,期待能学习。。。