Data Mining on Ice

Abstract

In an atmospheric neutrino analysis for IceCube’s 59-string configuration, the impact of detailed feature selection on the performance of machine learning algorithms has been investigated. Feature selection is guided by the principle of maximum relevance and minimum redundancy. A Random Forest was studied as an example of a more complex learner. Benchmarks were obtained using the simpler learners k-NN and Naive Bayes. Furthermore, a Random Forest was trained and tested in a 5-fold cross validation using 3.5 × 104 simulated signal and 3.5 × 104 simulated background events.