Abstract
We propose an approach to the handling of missing inputs in neural networks for constructing neural-network-based diagnosis systems from incomplete data with missing attribute values. In our approach, each unknown attribute value is represented by an interval which includes all the possible values of that attribute. Therefore the incomplete data with missing attribute values are transformed into interval data. Then, three learning algorithms for multi-class classification problems of interval input vectors are derived in a similar manner as the back-propagation algorithm. Four classification rules based on interval output vectors from trained neural networks are also proposed. The proposed approach is applied to the medical diagnosis of hepatic diseases. High performance of the neural-network-based diagnosis system constructed by our approach is demonstrated by comparing it with the existing fuzzy-rule-based diagnosis systems.
