Kaggle's Acquire Valued Shoppers Challenge

This repo holds the Python and R code I used to make submision to Kaggle's Acquire Valued Shoppers Challenge.

Method

• This competition is about dealing with "big data" (around 20G+) and mostly about feature engineering. To handling this amount of data and construct feature, I rely on Python. In specific, the corresponding code in the ./Python folder are based on Kaggler @Triskelion's implementation. (Many Thanks to @Triskelion)

• For the training phase, I have tried various models:

• Gradient Boosting Machine: gbm in R, xgboost, GradientBoostingClassifier in scikit-learn. Among these, I have little luck with scikit-learn's implementation. It scores 0.60535 on the public LB and 0.60176 on the private one. Other methods, even with small depth and small number of trees tend to overfit.

• Random Forest: randomForest in R, RandomForestClassifier and RandomForestRegressor in scikit-learn. For RandomForestRegressor, I regressed to the repeattrips variable instead of the repeater factor. RandomForestClassifier seems work best with score around 0.602.

• GLM: glmnet in R. This linear model turns out to work best with 0.60908 on the public LB and 0.59921 on the private one. Obvious, I have overfitted the public LB :-(

Requirement

• Python: numpy, pandas, scikit-learn

• R: glmnet, data.table, bit64, caTools, foreach

Instruction

• download data from the competition website and put all the data into ./Data dir

• ./Data/transactions.csv
• ./Data/trainHistory.csv
• ./Data/testHistory.csv
• ./Data/offers.csv

• put all the code into ./ dir:

• ./Python/...
• ./R/...

• run ./Python/generate_features.py to generate the features

• run ./Python/train_gbm.py to train GradientBoostingClassifier (scores 0.60535 on the public LB and 0.60176 on the private one)

• run ./R/train_glm.R to train unbagging/bagging version of glm (scores 0.60908 on the public LB and 0.59921 on the private one)