Implementing Streaming Parallel Decision Trees on Graphic Processing Units

Sammanfattning: Decision trees have long been a prevalent area within machine learning. With streaming data environments as well as large datasets becoming increasingly common, researchers have developed decision tree algorithms adapted to streaming data. One such algorithm is SPDT, which approaches the streaming data problem by making use of workers on a network combined with a dynamic histogram approximation of the data. There exist several implementations for decision trees on GPU, but those are uncommon in a streaming data setting. In this research, conducted at RISE SICS, the possibilities of accelerating the SPDT algorithm on GPU is investigated. An implementation is successfully created using the CUDA platform. The implementation uses a set number of data samples per layer to better fit the GPU platform. Experiments were conducted to investigate the impact on both accuracy and speed. It is found that the GPU implementation performs as well as the CPU implementation in terms of accuracy, suggesting that using small subsets of the data in each layer is sufficient for making accurate split decisions. The GPU implementation is found to be up to 113 times faster than the reference Scala CPU implementation for one of the tested datasets, and 13 times faster on average over all the tested datasets. Weak parts of the implementation are identified, and further improvements are suggested to increase both accuracy and runtime performance.