Using fuzzy support and confidence setting to mine interesting association rules

This paper proposes a fuzzy model to derive the appropriate minimum support and confidence thresholds for mining the association rules. The traditional data mining technologies of association rules usually base on user-defined minimum support and confidence values. The most important problem is how to select the appropriate minimum support and confidence to find frequent itemsets. A priori algorithm, the widely adopted approach, exploits the following property to derive the frequent itemset: if an itemset is frequent, so are all its subsets. That is, apriori algorithm generates itemsets in a level-wise manner where each candidate in the jth iteration is generated from previous frequent (j-1)-itemsets. A generated candidate can be further pruned if any subset of size j-1 is not a frequent itemset. A priori algorithm relies on the essential assumption that all itemsets have a uniform minimum support value, i.e., we assume that all items in the dataset have the same nature, e.g., all the items have the same sale price or the same salability condition in different time intervals or locations. However, the assumption may not comply with the rules embedded in the large databases. Concept hierarchy is helpful to solve the problem and the support and confidence thresholds should vary especially while we consider items at different conceptual abstractions. For example, turkey and pumpkin pie are seldom sold together. However, if we look at the transactions in the week before Thanksgiving, we may discover that most transactions contain turkey and pumpkin pie. It means that we should apply different support values to different time intervals. In this paper, we present a framework of multilevel association rules mining in the presence of fuzzy concept hierarchies that would derive a reasonable minimum support and confidence setting without losing potential interesting rules.