Separator Database and SPM Tree Framework for Mining Sequential Patterns Using Prefixspan with Pseudoprojection

• Main Author: Dhany, Saputra
• Format: Thesis
• Published: 2008
• Subjects: QA75 Electronic computers. Computer science
• Online Access: http://utpedia.utp.edu.my/1069/
• Institution: Universiti Teknologi Petronas

Sequential pattern mining is a new branch of data, mining science that solves inter-transaction pattern mining problems. Efficiency and scalability on mining complete set of patterns is the challenge of sequential pattern mining. A comprehensive performance study has been reported that PrefixSpan, one of the sequential pattern mining algorithms, outperforms GSP, SPADE, as well as FreeSpan in most cases, and PrefixSpan integrated with pseudoprojection technique is the fastest among those tested algorithms. Nevertheless, peudoprojection technique, which requires maintaining and visiting the in-memory sequenced database frequently until all patterns are found, consumes a considerable amount of memory space and induces the algorithm to undertake many redundant and unnecessary checks to this copy of original database into memory when the candidate patterns are examined. Moreover,improper management of intermediate databases may adversely affect the execution time and memory utilization. In the present work, Separator Database is proposed to improve PrefixSpan with pseudoprojection through early removal of uneconomical in-memory sequenced database, whilst SPM-Tree framework is proposed to build the intermediated databases. By means of procedures for building index set of longer patterns using Separator Database, some procedure in accordance to in-memory sequence database can be removed, thus most of the memory space can be released and some obliteration of redundant checks to in-memory sequence database reduce the executiont ime. By storing intermediated atabasesin to SPM-Tree Framework,the sequence database can be stored into memory and the index set may be built. Using Java as a case study, a series of experiment was conducted to select a suitable API class named Collections for this framework.The experimental results show that Separator Database always improves, exponentially in some cases, PrefixSpan with pseudoprojection. The results also show that in Java, A Arraylist is the most suitable choice for storing Object and Arraylnt list is the most suitablec choice for storing integer data. This novel approacho for integrating separator Database and Framework using these choices of Java collections outperforms with pseudoprojectionin terms of CPU performance and memory. Future research includes exploring the use of separator Database in PrefixSpan with pseudoprojection to improve mining generalized sequential patterns, particularly in handling mining constrained sequential patterns.