Scalable Machine Learning on Popular Analytic Languages with Parallel Data Summarization

Machine learning requires scalable processing. An important acceleration mechanism is data summarization, which is accurate for many models and whose summary requires a small amount of RAM. In this paper, we generalize a data summarization matrix to produce one or multiple summaries, which benefits a broader class of models, compared to previous work. Our solution works well in popular languages, like R and Python, on a shared-nothing architecture, the standard in big data analytics. We introduce an algorithm which computes machine learning models in three phases: Phase 0 pre-processes and transfers the data set to the parallel processing nodes; Phase 1 computes one or multiple data summaries in parallel and Phase 2 computes a model in one machine based on such data set summaries. A key innovation is evaluating a demanding vector-vector outer product in C++ code, in a simple function call from a high-level programming language. We show Phase 1 is fully parallel, requiring a simple barrier synchronization at the end. Phase 2 is a sequential bottleneck, but contributes very little to overall time. We present an experimental evaluation with a prototype in the R language, with our summarization algorithm programmed in C++. We first show R is faster and simpler than competing big data analytic systems computing the same models, including Spark (using MLlib, calling Scala functions) and a parallel DBMS (computing data summaries with SQL queries calling UDFs). We then show our parallel solution becomes better than single-node processing as data set size grows.