Plausible neural network: Theories and applications.
详细信息
文摘
Multivariate data analysis has attracted quite a large amount of research effort from disciplines such as statistics, data mining, and artificial intelligence. Multivariate data analysis problems include classification, clustering, pattern recognition, function approximation, rule discovery, etc. Due to technology innovations in data acquisition, computer storage, and networking, data collection becomes routine and inexpensive. As a result, datasets in various fields grow rapidly, which contain patterns and knowledge valuable for applications like fraud detection, disease diagnosis, and security profiling. However, to make use of those datasets, efficient data analysis tools are required to extract useful patterns and knowledge hidden inside them.;This dissertation research extends and applies Plausible Neural Network (PNN), a neural network model developed by Chen in 2002, to various multivariate data analysis implementations. PNN provides a simple and unified architecture applying to different data analysis problems. In addition, the PNN architecture integrates categorical and continuous attributes within the same framework and handles missing data soundly and efficiently. Thus, PNN is more general in dealing with real-life data analysis problems. The major contributions of this dissertation to the PNN theory include a definition of energy function that can be used to measure the quality of PNN clustering and a tuning algorithm which improves the PNN classification performance. A software package, called PNNSolution, is implemented to facilitate the PNN application of general-purpose multivariate data clustering and classification. PNNSolution has been tested on many datasets from the UCI machine-learning repository. The preliminary test results are very promising and also give clues to further enhancing the PNN architecture.;Specifically, this dissertation research applies PNN to fuzzy membership function elicitation which bridges the gap of the statistic characteristics and the semantic meanings of the fuzzy sets. This dissertation also implements PNN function approximation. The implementation shows the superiority of PNN since PNN performs faster training and one-to-multiple approximation. Furthermore, this dissertation proposes a PNN framework for bioinformatics sequence mining and tests it on the exon/intron boundary prediction problem. The promising results show the capability of handling the large dataset in PNN implementation.