Performance Evaluation of Sequential Rule Mining Algorithms

Abstract

Data mining techniques are useful in discovering hidden knowledge from large databases. One of its common techniques is sequential rule mining. A sequential rule (SR) helps in finding all sequential rules that achieved support and confidence threshold for help in prediction. It is an alternative to sequential pattern mining in that it takes the probability of the following patterns into account. In this paper, we address the preferable utilization of sequential rule mining algorithms by applying them to databases with different features for improving the efficiency in different fields of application. The three compared algorithms are the TRuleGrowth algorithm, which is an extension sequential rule algorithm of RuleGrowth; the top-k non-redundant sequential rules algorithm (TNS); and a non-redundant dynamic bit vector (NRD-DBV). The analysis compares the three algorithms regarding the run time, the number of produced rules, and the used memory to nominate which of them is best suited in prediction. Additionally, it explores the most suitable applications for each algorithm to improve the efficiency. The experimental results proved that the performance of the algorithms appears related to the dataset characteristics. It has been demonstrated that altering the window size constraint, determining the number of created rules, or changing the value of the minSup threshold can reduce execution time and control the number of valid rules generated.