Contributions to parallel and distributed computing in knowledge discovery and data mining
Lozano Inca, Elio
CollegeCollege of Engineering
DepartmentDepartment of Electrical and Computer Engineering
MetadataShow full item record
Recently databases are increasing continuously without bound, due to new data acquisition technologies. One challenge is how to gain knowledge from these large data sets. In this thesis, we analyze and improve the algorithmic solution of four problems related to knowledge discovery and data mining, making use of parallel computing; we also compare our results with related works. We design two parallel algorithms for outlier detection; the first one is for finding distance-based outliers based on nested loops along with randomization and the use of a pruning rule. The second parallel algorithm is for detecting density-based local outliers. In both cases data parallelism is used. The star coordinates plot is a useful visualization technique, but it has some drawbacks. We enhance the traditional star coordinates plot introducing new parameters that will allow us to visualize the data points in two dimensions as polygons and in three dimensions as polyhedrons. In order to visualize large data sets and reduce its computational time, a parallel algorithm is also designed. We design a new meta-classifier algorithm, and its performance is compared with base classifier algorithms and Bagged based meta-classifier algorithms. Our meta-classifier algorithm gives better results compared to other meta-classifier algorithms. For speeding up its computation time as well as making it suitable for large data sets a parallel algorithm is developed. We develop a meta-clustering algorithm and compare its performance with two Bagged based meta-clustering algorithms, and hypergraph partitioning meta-clustering algorithm. Our proposed meta-clustering algorithm gives results close to the best clustering algorithm, and is more robust to the data dependency problem. A parallel algorithm to compute four meta-clustering algorithm is also designed. The experimental results of our collection of sequential and parallel programs is tested in two different clusters of Linux-based workstations using real-world databases available in the Machine Learning Repository of the University of California at Irvine.