An Application Of Multi-GPU-Based Association Mining Algorithm In Medical Data

With the development of Data Mining technology, Data Mining method has been widely used in more and more areas. In the medical field, by analyzing large medical data, we can find out the regularity and relationships between different diseases and other factors, which is significant to personalized treatment, predictive models of disease and other clinical trials analysis. In recent years, researchers both in domestic and foreign have done a lot of research in medical large data. They brought out many effective and applicable algorithms on classification, clustering and associations rules mining, which promote the development of large medical data industry.In these algorithms, Apriori algorithm has become the most widely used algorithm in association rules mining of serious diseases such as lung cancer, diabetes and cardiovascular disease for its high accuracy and easy operation. However, with the increasing amount and the increasing operation complexity of medical data, the running time of Apriori algorithm is getting longer and longer, which restricts the practical application of Apriori. Therefore, how to use Apriori to mine association rules over massive data under the big data era is a serious problem. Google proposed MapReduce to accelerate Apriori algorithm effectively by using high performance computing cluster, which is the most widely used method currently. However, many researchers and small and medium enterprises cannot afford the high cost of constructing cluster. Compared with the CPU, GPU (Graphic Processing Unit) is very suitable to solve the problem of massive data parallel computing for its larger memory bandwidth, more execution units and the lower cost. Therefore, we present a multi-GPU-based parallel Apriori algorithm in this paper. We have proved that our method can achieve a good acceleration effect with a much lower cost, which makes it possible to be used widely.In this paper, we parallelize the Apriori algorithm. We transfer the calculation of support count from CPU to GPUs by assigning one transaction to each thread to taking the advantages of the parallel computing capabilities of GPUs. According to the data test, the multi-GPU-based parallel Apriori algorithm proposed in this paper can not only greatly reduce the time cost, but also have a great scalability of GPU.