Transaction-filtering data mining and a predictive model for intelligent data management

ChenHan Liao, AMAC Group, School of Engineering, Cranfield University i This thesis, first of all, proposes a new data mining paradigm (transaction-filtering association rule mining) addressing a time consumption issue caused by the repeated scans of original transaction databases in conventional associate rule mining algorithms. An in-memory transaction filter is designed to discard those infrequent items in the pruning steps. This filter is a data structure to be updated at the end of each iteration. The results based on an IBM benchmark show that an execution time reduction of 10% 19% is achieved compared with the base case. Next, a data mining-based predictive model is then established contributing to intelligent data management within the context of Centre for Grid Computing. The capability of discovering unseen rules, patterns and correlations enables data mining techniques favourable in areas where massive amounts of data are generated. The past behaviours of two typical scenarios (network file systems and Data Grids) have been analyzed to build the model. The future popularity of files can be forecasted with an accuracy of 90% by deploying the above predictor based on the given real system traces. A further step towards intelligent policy design is achieved by analyzing the prediction results of files’ future popularity. The real system trace-based simulations have shown improvements of 2-4 times in terms of data response time in network file system scenario and 24% mean job time reduction in Data Grids compared with conventional cases.