A parallel approach for decision trees learning from big data streams

Abstract

In this paper we introduce PdsCART, a parallel decision tree learning algorithm. There are three characteristics that are important to emphasize and make this algorithm particularly interesting. Firstly, the algorithm we present here can work with streaming data, i.e. one pass over data is sufficient to construct the tree. Secondly, the algorithm is able to process in parallel a larger amount of data stream records and can therefor handle efficiently very large data sets. And thirdly, the algorithm can be implemented in the MapReduce framework. Details about the algorithm and some basic performance results are presented.