shubhampachori12110095 / amazon Goto Github PK

This code produces a portion of our 1st place code from the Kaggle Amazon Access Competition.

My partner, Paul Duan, produced the other portion. At the time we also had code to blend our various model outputs.

see: https://www.kaggle.com/c/amazon-employee-access-challenge

and: https://www.kaggle.com/c/amazon-employee-access-challenge/leaderboard/private

Also included is an ipython notebook that uses the same data. I used this notebook to present a practical introduction to scitkit learn and the random forest algorithm to the Pittsburgh Python User Meetup group.

About the Data: The goal was to predict employee resource access grants from categorical job description data. The scoring metric is AUC (area under the ROC curve). There are only 9 categorical input feature columns, one of which is completely redundant. There are roughly 30,000 training rows and 50,000 test rows.

About the Code: The general strategy was to produce 2 feature sets: one categorical to be modeled with decision tree based approaches and the second a sparse matrix of binary features, created by binarizing all categorical values and 2nd and 3rd order combinations of categorical values. The latter features could be modeled with Logistic Regressoin, SVMs, etc. The starting point of this latter set of code was provided on the forums by Miroslaw Horbal. The most critical modeification I made to it was in merging the most rarely occuring binary features into a much smaller number of features holding these rare values.

ensemble.py generates the categorical featuers and runs the ensemble of tree models. logistic_features.py generates the binarized categorical data. It's currently set up to predict using a NaiveBayes algorithm (for speed), but a 1-line change could replace that with Logistic regression, which should be much higher scoring.

It was run most recently on windows 8 with python 2.7.5 and the following package versions:

scikit-learn 14.1

pandas 12.0

numpy 1.8