Parallel Implantation of Frequent Itemset Mining Using Inverted Matrix Based on OpenCL

Abstract

Extracting knowledge in the form of frequent itemsets and association rules deserves great importance in the field of data mining. Apriori algorithm suffers from multiple scans of the database and thus forms high memory dependency. On the other hand frequent pattern tree (FP tree) growth algorithm becomes impractical for large databases due to memory-based data structure. An efficient approach of inverted matrix with COFI (co-occurrence frequent item) tree alleviates disadvantages of both the above-mentioned algorithms. For massively large computations, modern GPUs provide a large set of parallel processors which facilitate in general-purpose computing. General purpose graphical processing unit (GPGPU) is way of utilizing the existing GPU for general purpose use. Open computing language (OpenCL) provides a standard for cross-platform programming on modern processors such as many-core CPUs and GPUs. As inverted matrix approach is advantageous over other algorithms, it is desirable to form it parallel to OpenCL. We have proposed a new technique called CLInverted matrix itemset mining, which is an advancement over existing techniques and contributes to load sharing. The proposed architecture in this paper highlights the inverted matrix approach implantation based on OpenCL framework. In experiments we have compared the results of serial and parallel versions of the proposed approach on various OpenCL devices.