Feature Engineering on Rainfall Matrices

We apply traditional machine learning techniques to the problem of predicting expected rainfall. In terms of feature engineering where each observation is a real valued matrix, we apply the TF-IDF to SVD transformation popular in natural language processing. Given each observation is a matrix, we consider Pearson correlation between the columns of each observation matrix. Furthermore, we construct meta features based upon a gradient boosting model trained on the training data set's outliers. When these methods are used in conjunction with a bagged gradient boosting model, we scored 11th place on Kaggle's How Much Did It Rain? II challenge.