Cornell Research Proposal

Research Starting Plan

Jinning Li, 07/24/2018

Code Github: https://github.com/jinningli/ad-placement-pytorch

01 Code review of CrowdAI Rank1

  • Use a FTRL model https://github.com/alexeygrigorev/nips-ad-placement-challenge
  • Mainly Feature Processing
  • delete the first two features
  • replace any feature larger than 1 as 1
  • process the result using sigmoid function and some other tricks.

[FINISH]

02 Try some mainstream linear models

  • Simple LR [FINISH]
  • XGBoost
  • LightGBM
  • FFM

Also try to adapt the objective function with propensity value.

03 Try some Deep Learning models

  • CNN
  • Deep FFM > https://arxiv.org/abs/1703.04247
  • DCN > https://arxiv.org/abs/1708.05123
  • Neural FM > https://arxiv.org/abs/1708.05027
  • Wide & Deep > https://arxiv.org/abs/1606.07792
  • PNN > https://arxiv.org/abs/1611.00144
  • Deep Interest > https://arxiv.org/abs/1706.06978

Consider how to combine propensity value

04 My own model

Develop my own propensity weighted deep click model

05 Apply some feature tricks

  • repeated features
  • feature overlap