This is the code for Kaggle - Don't Get Kicked! The problem is to predict if a car purchased is a kick, which means the car purchased by an auto dealership at an auto auction might have serious issues that prevent it from being sold to customers. There are about 30 features of a car given, most of which are categorical features like the model of the car, which country produced the car etc.. Also, this is a 
skewed class problem as only 1/7 of the cars are marked as kicks. For more details, check Kaggle’s official description http://www.kaggle.com/c/DontGetKicked.
My approach is based on random forest and most of my time was spent on feature engineering. For numerical features, I found differences between prices quite informative, and I added about 14 features based on these. For categorical features, I binarized all of them except "model" and "submodel". Also for each categorical feature, I added the log-likelihood ratio of it, which boost the result a little bit. Totally I got 500+ features, which turned out to be  redundant. So I trained a decision tree with these 500+ features and selected best 300 features based on Gini impurity. Since this is a skewed classes problem, I down sampled the training set to make pos/neg ratio 1/1.5.
With about 1h training, my best performance is 0.25147, ranking 31 among 571 groups. And the leader is 0.26720.