hengqujushi / nips-ad-placement-challenge

The winning solution to the Ad Placement Challenge (NIPS'17 Causal Inference and Machine Learning Workshop)

Ad Placement Challenge

The winning solution to the Ad Placement Challenge by Criteo for the NIPS'17 Causal Inference and Machine Learning Workshop

• Report: https://arxiv.org/abs/1712.01913 (also in report/)
• Slides: https://www.slideshare.net/AlexeyGrigorev/ad-placement-challenge
• Video: https://www.youtube.com/watch?v=d5I0baV51vk

The solution is simple:

• Take the data as is, train a FTRL model (using libftrl-python)
• Post-process the predictions:
  • Apply the sigmoid fuction to the predictions and scale the result by some constant value
  • For each group add +15 to the max value

Environment:

• ubuntu 14.04
• 8 cores, 32 gb of ram
• anaconda with python 3.5.2
• additional packages not from anaconda: tqdm and ftrl

Running the solution

• mkdir tmp
• python 01-cv-split.py # takes ~22 mins
• python 02-prepare-data.py # ~17 mins
• python 03-train-model.py # ~25 mins
• python 04-predict.py # ~4.5 hours
• gzip pred_ftrl.txt
• submit pred_ftrl.txt.gz (~1.5 hours)