Combining GPU and Apache Spark for Evolving Explainable Decision Trees on Large-Scale Data

Decision trees (DTs) belong to explainable machine learning techniques applicable for classification and regression problems. Traditionally, DTs are built through a top-down greedy search, which is usually fast but may result in sub-optimal solutions. One of the alternatives is the use of evolutionary algorithms (EAs), allowing for more global exploration that can yield simpler and accurate DTs. However, the EA-based DT induction is computationally expensive, especially for large-scale data. Therefore, different parallel and distributed accelerators are continuously explored to alleviate its high computing requirements. In this paper, we propose a hybrid Spark+GPU supported solution on commodity hardware. The Spark-based component distributes the computations over a cluster, while inside each cluster node, they are further parallelized on a GPU. Experimental validation on various datasets shows that the hybrid solution significantly accelerates the EA-based DT induction and scales well by simply adding more commodity nodes. Moreover, it is much more price-effective than a multi-GPU supported one running on costly HPC servers.