Parallel Frequent Pattern Mining without Candidate Generation on GPUs

The graphics processing unit (GPU) has evolved into a key part of today's heterogeneous parallel computing architecture. A number of influential data mining algorithms have been parallelized on GPUs including frequent pattern mining algorithms, such as Apriori. Unfortunately, due to two major challenges, the more effective method for mining frequent patterns without candidate generation named FP-Growth has not been implemented on GPUs. Firstly, it is very hard to efficiently build the FP-Tree in parallel on GPUs as it is an inherently sequential process. Secondly, mining the FP-Tree in parallel is also a difficult task. In this paper, we propose a fully parallel method to build the FP-Tree on CUDA-enabled GPUs and implement a novel parallel algorithm for mining all frequent patterns using the latest CUDA Dynamic Parallelism techniques. We show that, on a range of representative benchmark datasets, the proposed GPU-based FP-Growth algorithm can achieve significant speedups compared to the original algorithm.