Study On Algorithms For Frequent Patterns Mining Based On Memory Indexing

In this thesis, we present a pattern growth algorithm for frequent patterns mining (called MIndexing algorithm). We adopt an indexing structure to grow the pattern length in a projected database, which may reduce the CPU time and save the memory consuming. The indexing structure is a prefix tree structure, which is also a compressed structure for current frequent patterns. We test our algorithm versus other algorithms on real world datasets and IBM artificial datasets. The empirical results illustrate that the MIndexing algorithm performs better than Apriori and FP-growth methods when processing sparse data datasets that may contain long patterns.While mining frequent patterns, we usually want to get different sets of itemsets based on different support levels to achieve useful information. It may waste a lot of time if we mine from scratch every time. In this thesis, we present an efficient approach to reduce the executing time of multiple mining processes. After MIndexing finds all frequent patterns of a specified smallest support level, the subsequent mining processes of other support levels become relatively simple and easy in our approach. The empirical results show that MIndexing outperforms Apriori and FP-growth methods in the same conditions, and achieves a great performance in the multiple searching. Event prediction is one of the main purposes in association rule mining.In this thesis we propose a method for the efficient prediction of rare events, i.e., engine failures, stock price situations and market analysis etc. We associate transaction as ordinal event series that occur in equal length intervals, such as one day, one week or even one month etc. We formulate the problem of rare events prediction and use MIndexing to mine valuable patterns for the fast and accurate prediction of user-specified rare events.