A Review and Analysis of a Parallel Approach for Decision Tree Learning from Large Data Streams

TL;DR

This paper presents pdsCART, a MapReduce-enabled parallelization of dsCART for streaming data. By distributing record processing horizontally and aggregating statistics via local and global histograms, pdsCART achieves near-identical tree structures to dsCART with significantly reduced training time on large data streams. The work demonstrates strong scalability using synthetic and real streaming datasets, and analyzes how factors like feature count and histogram binning affect performance. The approach offers practical benefits for real-time, scalable decision tree learning in distributed environments, with future work targeting further scalability and parameter analysis. All mathematical notation is presented in $...$ format, including the threshold parameter $\Theta$ and the data distribution $R/P$.

Abstract

This work studies one of the parallel decision tree learning algorithms, pdsCART, designed for scalable and efficient data analysis. The method incorporates three core capabilities. First, it supports real-time learning from data streams, allowing trees to be constructed incrementally. Second, it enables parallel processing of high-volume streaming data, making it well-suited for large-scale applications. Third, the algorithm integrates seamlessly into the MapReduce framework, ensuring compatibility with distributed computing environments. In what follows, we present the algorithm's key components along with results highlighting its performance and scalability.