Measures of similarity, relevance and expertise in case-based reasoning systems

The thesis proposes an information-theoretic approach to fundamental aspects of Case-Based Reasoning (CBR) over a database involving three-dimensional medical images. Given information from a new case, a CBR system needs to retrieve records of cases which are similar to the new case. This thesis proposes a mutual information approach to measuring similarity. It extends the transformational approach to similarity by introducing the concept of relevant information. The thesis demonstrates how to apply these approaches to images. The thesis also considers issues concerning the extent to which the recommendations of a CBR system can be trusted. The expertise of a CBR system is that part of the problem domain for which the system can return trustworthy results. The thesis also proposes efficient techniques for searching and ranking similar items from a database, particularly when the database is distributed geographically across a number of independent cooperating institutions.