| There exist many databases containing unprocessed raw data. However, there are few tools to extract information from such databases. With increasing size and number of databases, there seems to be no shortage of data or examples; if anything, the problem is what to do with them. In general, relationships and regularities in the data can be considered as "knowledge". The process of revealing these regularities constitutes "knowledge extraction". Emerging Neural Network techniques are promising tools for automated knowledge extraction. The objective of this study is to perform automated knowledge extraction from a raw database of records (e.g., hospital patient records) where each record consists of several fields or features (e.g., patient age, sex, etc.). There are two different kinds of knowledge that we would like to extract from such a database. First, we are interested in extracting unknown functional dependencies that may exist between several features (e.g., patient age and blood pressure). This problem is known in statistics as nonparametric regression analysis. In this study, we introduce a new neural network algorithm for nonparametric regression, based on the Kohonen self-organizing maps. Experimental results show the superiority of this technique over the existing statistical approaches. The second problem of knowledge extraction addressed in this thesis is the ability to perform automated inferencing (prediction, diagnosis). For example, we seek to diagnose a patient using available information such as age, sex, symptoms and test results, as well as using a database of previously diagnosed patients. Here, the problem is to combine and relate dissimilar features (e.g., sex, age, blood pressure) in a meaningful fashion. These features may vary in type, range of values, and also may have hidden statistical dependencies between them. Most existing neural approaches rely on the back propagation network (or its variants) to perform knowledge extraction from raw data. In contrast, we argue that for real-life applications the issue of data representation (preprocessing, encoding) is more important than the choice of a neural network. This study introduces a new paradigm for neural network-based diagnostic systems that emphasize informative data representation and encoding. |
