Criteo/Kaggle Competition of CTR prediction
This is the code was written for the Kaggle Criteo Competition of CTR prediction.
Since the data are highly sparse, the basic methodology is to use logistic regression with appropriate quadratic/polynomial feature generation and regularization to make sophisticated and over-fitting-tractable models. Vowpal Wabbit is the major machine learning software used for this project. Since the data size is challenging in terms of my personal workstation (a single quad-core CPU), the techniques of feature selection and model training are selected based on the trade off between performance and CPU/RAM resource limit.
Please note that the code was written for my personal learning and practice in new features of Java 8 and Python 3.4 in Ubuntu 14.04. The code cannot be run in early versions of these two languages or other OSs. Compatibility is not considered here.
$ cd display-ad-java $ mvn package # or mvn install
$ cd work $ ../run.sh