Dynamically creating indices for two million cases: a real world problem

Efficiently indexing and retrieving cases from a very large case library are major concerns when building a Case-Based Reasoning (CBR) system. Most CBR research has focused on representation of cases, how to identify features that should be used for retrieval; and similarity measurement between values of attributes. In this paper, we propose a method for dynamically creating indices, and, also different similarity-measurement methods for different types of attributes. We also discuss the use of a relational database for representing cases, taxonomy knowledge, and spatial information. Our real world problem domain consists of 2 million incomplete insurance eases, with 30 different attributes. Even though all of these are valid cases, only 10 percent of these policies have lodged claims. These situations create a very complex case base for reasoning and problem solving. In response to this complexity, the approach adopted in building our CBR system involves a considerable amount of statistical pre-analysis of the contents of the case base to generate domain knowledge that could be used by the “Dynamic Index Creation Mechanism”. The main contribution of this paper is in describing the techniques used in our CBR system to dynamically create indices for the purpose of effective case retrieval.