| dc.description.abstract | The main focus of this report is on frequent intra- and inter-transaction itemset mining, specifically regarding parallel algorithms on shared memory multi-core processor systems. Multiple state-of-the-art algorithms are presented for both type of frequent itemset mining problems. Three novel parallel algorithms are presented and implemented, these include a parallel intra and two parallel inter algorithms. In addition to this, three state-of-the-art intra algorithms are implemented as well.
A thorough experimental procedure is conducted in order to analyse the implemented methods in-depth. This procedure includes the creation of multiple synthetic datasets with various characteristics. All intra and inter algorithms are tested on these datasets, in order to see how the behaviour of the algorithms changes with different dataset attributes. Multiple real datasets, which include already existing and created datasets, are used during the experiments as well. All these tests include a comparison between the algorithms, with different algorithm parameters.
The results from the experiments show that the utilization of multi-core processors on frequent itemset mining methods is good. In the best cases, on the tested system, the speedup was almost 35 times better than single threaded execution. However, the results also show that whenever a large quantity of frequent itemsets are generated, the system bus becomes the bottleneck. All in all the viability of parallelization of frequent intra- and inter-transaction itemset mining algorithms is advantageous. | |
