A parallel algorithm to induce decision trees for large datasets

This paper introduces a new parallel algorithm called ParDTLT and discusses some of its advantages with respect to a set of well known sequential and parallel algorithms. The parallel process occurs in every node in the decision tree, which is constructed during the supervised training phase. The basis of the distribution of a parallel task is on the attributes of the training objects and the growing of the tree is based on two criteria, who are defined by the maximum number of training objects that every node can support and an entropic gain ratio criterion. Different experiments have been made to compare the behavior of the parallel algorithm ParDTLT with the behavior of the sequential algorithms C4.5, VFDT, YaDT and DTLT and with the parallel algorithm called Synchronous. The experimental results show that ParDTLT keeps the quality of classification and it reduces the execution time.