Dp 518 en
Distributed computation of inductive models
Author: Jakub Špirk
Development of good predictive regression models and classifiers is computationally very expensive. This thesis aims to present changes made to FAKE GAME software in order to enable distributive computation of models it generates. I propose two designs and implementations of necessary adjustments to the software; the first one changes the current process of model development to support multi-threading, the later is integration with Storm framework and distributive file system HDFS to support parallel execution across cluster systems. Experiments performed on both real and artificial datasets confirm considerable speedup of development process for most real models as well as scalability of application on clusters.