Fitness evaluation reuse for accelerating GPU-based evolutionary induction of decision trees

摘要

Decision trees (DTs) are one of the most popular white-box machine-learning techniques. Traditionally, DTs are induced using a top-down greedy search that may lead to sub-optimal solutions. One of the emerging alternatives is an evolutionary induction inspired by the biological evolution. It searches for the tree structure and tests simultaneously, which results in less complex DTs with at least comparable prediction performance. However, the evolutionary search is computationally expensive, and its effective application to big data mining needs algorithmic and technological progress. In this paper, noting that many trees or their parts reappear during the evolution, we propose a reuse strategy. A fixed number of recently processed individuals (DTs) is stored in a so-called repository. A part of the repository entry (related to fitness calculations) is maintained on a CPU side to limit CPU/GPU memory transactions. The rest of the repository entry (tree structures) is located on a GPU side to speed up searching for similar DTs. As the most time-demanding task of the induction is the DTs’ evaluation, the GPU first searches similar DTs in the repository for reuse. If it fails, the GPU has to evaluate DT from the ground up. Large artificial and real-life datasets and various repository strategies are tested. Results show that the concept of reusing information from previous generations can accelerate the original GPU-based solution further. It is especially visible for large-scale data. To give an idea of the overall acceleration scale, the proposed solution can process even billions of objects in a few hours on a single GPU workstation.