AN EMPIRICAL STUDY OF CONTEXT BASED SEQUENTIAL PATTERN MINING ALGORITHMS EFFICIENCY

摘要

Methods of patterns detection in the sets of data are useful and demanded tools in a knowledge discovery process. The problem of searching patterns in set of sequences is named Sequential Patterns Mining. It can be defined as a way of finding frequent subsequences in the sequences database. The patterns selection procedure may be simply understood. Every subsequence must be enclosed in the required number of sequences from the database at least to become a pattern. The number of a pattern enclosing sequences is called a pattern support. The process of finding patterns may look trivial but its efficient solution is not. The efficiency plays a crucial role if the required support is lowered. The number of mined patterns may grow exponentially. Moreover, the situation may change if the problem of Sequential Patterns Mining will be extended further. In the classic definition the sequence is a list of ordered elements containing only non-empty sets of items. The Context Based Sequential Patterns Mining adds uniform and multi-attribute contexts (vectors) to the elements of the sequence and the sequence itself. Introducing contexts significantly enlarges the problem search space. However, it brings some additional occasions to constrain the mining process, too. This enhancement requires new algorithms. Traditional ones are not able to cope with non-nominal data directly. Algorithms derived straightly from traditional algorithms were verified to be inefficient. This study evaluates efficiency of novel ContextMapping and ContextMappingHeuristic algorithms. These innovative algorithms are designed to solve the problem of Context Based Sequential Pattern Mining. This study answers in what scope the algorithms parameterization impacts on mining costs and accuracy. It also refers the modified problem to the traditional one pointing at the common and uncommon properties and drawing perspective for further research.