On Efficient Mining of Frequent Itemsets from Big Uncertain Databases

Abstract

In the current era of information, communication, and technology the data is being generated at an exponential rate. This provides machine learning and data mining algorithms an opportunity to learn from huge data repositories. However, at the same time, the big data poses many challenges. Data uncertainty being the key concern of the modern data mining systems. This work addresses the problem of extracting frequent itemsets from such large uncertain databases to assist the decision makers in understanding the non-trivial data trends. The usual technique utilized to find frequent itemsets from uncertain databases is known as the Possible Word Semantics (PWS). However, as the database size increases, PWS suffers from performance issues. Therefore, there is a need for efficient frequent pattern mining algorithms. This work presents three techniques to address the issue at hand, namely: 3D linked array-based strategy, connected tree technique, and average probability-based setup with the support of a tree data structure. The objective here is to minimize computational cost by traversing the database only once. The 3D linked array-based solution scans the database only once and stores the support information of the item and its association with other items within the 3D array. For the tree-based method, 1D array is associated with each node of the tree, comprising of support information of the database items and their associations with other items. The average probability-based approach computes the average probability factor and utilizes it to map the uncertain database to a tree. The current proposal addresses attribute uncertainty as well as the tuple uncertainty to map large uncertain databases to the proposed data structures. In addition to introducing the three data structures, this work also presents algorithms to extract frequent itemsets. The proposal is compared with four recent works done in this domain for uncertain data, namely, mining threshold-based (MB) technique, frequent itemsets using nodesets (FIN), prepost + , and uncertain apriori (UApriori). Experiments are performed utilizing four benchmark datasets. The results obtained suggest better performance of the three techniques presented here, while consuming 60% less execution time.