Towards Lifetime Maintenance of Case Based Indexes for Continual Case Based Reasoning

One of the key areas of case based reasoning is how to maintain the domain knowledge in the face of a changing environment. During case retrieval, a key process of CBR, feature-value pairs attached to the cases are used to rank the cases for the user. Different feature-value pairs may have different importance measures in this process, often represented by feature weights attached to the cases. How to maintain the weights so that they are up to date and current is one of the key factors determining the success of CBR. Our focus in this paper is on the lifetime maintenance of the feature-weights in a case base. Our task is to design a CBR maintenance system that not only learns a user's preference in the selection of cases but also tracks the user's evolving preferences in the cases. Our approach is to maintain feature weighting in a dynamic context through an integration with a learning system inspired by a back-propagation neural network. In this paper we explain the new system architecture and reasoning algorithms, contrasting our approach with the previous ones. The effectiveness of the system is demonstrated through experiments in a real world application domain.