Abstract :Over the last two decades the amount of data collected by enterprises has grown to tremendous levels. As the amount of data has grown, the need to develop techniques that process that data quickly and efficiently has also grown. One of these techniques, data mining, consists in finding patterns in large sets of data. Association rule mining, also known as frequent itemset mining, is a descriptive data mining technique which uses specific probabilities, especially those known as support and confidence, to find significant associations among items in transactional data. An association rule is a conditional statement that gives predictions on the occurrence of an itemset given the occurrence of other items in a particular transaction set. These kinds of rules are often used for basket analysis in marketing applications. The associations are converted into rules, if their probabilities are above the thresholds determined at the beginning of the analysis. The analysis can produce far more rules than can be easily managed. Thus, a common problem for association rule mining is that the number of association rules found are still too large to be of use to decision makers. This largely stems from the fact that association rule mining generates many redundant rules, but it is also because the rules are generated in a logically simple form which makes it impossible to multiple rules into a single rule. In this paper we propose an algorithm and a heuristic technique for simplifying the rules generated by association rule mining algorithms. Our algorithm first introduces a very simple and effective method of eliminating redundant rules. We then show that by rewriting the rules in a higher order logic, we can often further reduce the rules. Student preferences in the elective courses in a faculty of business and sample public data on supermarket baskets are used as sample applications of our method. In each case, the association rules are drastically simplified and it is easy to see that these results will generalization to other data sets. The FP-Growth algorithm is used as a starting point to extract the frequent itemsets based on support values. Association rules are then obtained through confidence values calculated over the frequent items sets. Additional metrics such as lift and p-s ratio are also used to detect the significance of the associations. Although we apply the FP-Growth algorithm to extract the rules, the proposed approach for reducing the number of rules is independent from the algorithm that is used to obtain associations. Keywords : Association Rule Mining, Rule Reduction, Heuristic Algorithm
