Thesis Type: Postgraduate
Institution Of The Thesis: Orta Doğu Teknik Üniversitesi, Faculty of Engineering, Department of Industrial Engineering, Turkey
Approval Date: 2016
Student: SANAM AZADİAMİN
Supervisor: GÜLSER KÖKSALAbstract:
Association rules are an important set of data mining results, which are helpful in handling large amount of data and extracting useful association information from them. There are many algorithms developed for finding interesting association rules and also some other algorithms for rule reduction purposes. All of the proposed methods have some strong and weak points, which can be useful according to their application areas. In the literature, there exist several comparison studies trying to find the best algorithm according to the user’s interests. But every comparison approach considers these algorithms using different measures, and it is hard to assess performance of an algorithm with respect to a measure since interesting association rules are unknown. A novel comparison method has been proposed by Jabarnejad (2010) based on interesting rules generated by logistic regression to compare rule reduction algorithms. In this study, this approach is extended to cover all association rule mining algorithms, on a broader set of test data developing and using relevant vi comparison measures. This approach utilizes design and analysis of experiments to generate test data. Furthermore, it defines several comparison measures, and the dependency and importance of these measures are analyzed using statistical methods such as factor analysis, ANOVA and nonparametric hypothesis tests. Finally, if statistical analyses show significant differences between applied association rule mining methods, it handles multiple comparisons using PROMETHEE. The approach is demonstrated by comparing three association rule mining algorithms. The results are discussed and future research directions are presented.