Multi-GPU approach to global induction of classification trees for large-scale data mining

Abstract This paper concerns the evolutionary induction of decision trees (DT) for large-scale data. Such a global approach is one alternatives to top-down inducers. It searches tree structure and tests simultaneously thus gives improvements in prediction size resulting classifiers many situations. However, it population-based iterative that can be too computationally demanding apply big data mining directly. The demonstrates this barrier overcome by smart distributed/parallel processing. Moreover, we ask question whether truly compete with greedy systems For purpose, propose novel multi-GPU approach. incorporates knowledge DT algorithm parallelization together efficient utilization memory computing GPU’s resources. are performed on CPU, while fitness calculations delegated GPUs. Data-parallel decomposition strategy CUDA framework applied. Experimental validation both artificial real-life datasets. In cases, obtained acceleration very satisfactory. solution able process even billions instances few hours single workstation equipped 4 impact characteristics (size dimension) convergence speedup search also shown. When number GPUs grows, nearly linear scalability observed what suggests boundaries fading.