Clustering algorithms, classification algorithms and their applications in medical databases.

摘要

Data mining is a process of discovering hidden patterns and relationships in large databases using various techniques such as clustering and classification. Clustering is the process of discovering groups of data, such that the intra-cluster similarity is maximized and the inter-cluster similarity is minimized. Many of the partitional clustering algorithms such as PAM, CLARA and CLARANS have failed to identify natural clusters of arbitrary shapes and sizes. It is required by these algorithms to provide the number of clusters in advance, which is very difficult to identify for high dimensional, large data sets. Hierarchal clustering algorithms such as CURE and ROCK have been developed which have overcome the limitations of partitional algorithms. However, these algorithms are based on a static model and hence they have failed to discover natural clusters. CHAMELEON is a hierarchal clustering algorithm that measures the similarity of two clusters based on a dynamic model. In the clustering process, two clusters are merged only if the relative closeness and relative inter-connectivity between the two clusters are greater than the threshold values. This property ensures that the natural clusters with arbitrary shapes, sizes and densities are identified.; I have implemented CHAMELEON in Maple 8.0 using the C library available in METIS and hMETIS graph partitioning packages and the mining software developed by Dr. Quoc-Nam Tran, Associate Professor, Lamar University. I have achieved a considerable improvement of 35% in the time for executing the program on the benchmark data sets.; Classification, also called supervised clustering, is another data mining technique. In this technique, a model is constructed using a training data set, which is then tested and used for classifying records whose class labels are unknown. I have implemented classification using Gini and Entropy based approaches and applied the program on thrombosis data sets. I have also compared the results of both the approaches and identified interesting rules for classifying thrombosis into any of the four possible categories namely, type 0, type 1, type 2 and type 3.