Application of Neural Networks for the Extraction & Characterization of Knowledge Contained in Databases

Abstract : The ability to automatically discover relationships contained within data, quantify their strength, and present them graphically to the user for visualization is defined as Relationship Discovery . This capability was the major research effort during Phase I of this SBIR Project. The detection of relationships is a necessary precursor to the modeling step where the detected relationships are modeled using powerful neural network paradigm capable of capturing the nonlinear relationships in the data. As a result of this Phase I SBIR Project, a relationship discovery capability has been developed to automatically determine the existence of a relationship in each sub-space and to determine the strength of the relationship. This capability is then used to produce a comprehensive listing of variables that are related to other variables and the strength of the relationship. As a result of the development and testing of these concepts on data sets, several necessary enhancements to the overall Data Base Mining system concept were identified. These are Missing Value Prediction, Bad Data Detection, and Data Redundancy.