Case retrieval nets as a model for building flexible information systems (Thesis, Humboldt-Univ. Berlin, 1999) (Q2726246)

Mario Lenz introduces case retrieval nets as a new method to organize the memory of case-based systems. A case retrieval net is inspired by neural networks and associative memory. A case is represented by a set of Information Entities (IEs)describing its features/attributes. IEs are the atomic pieces of knowlege one wants to represent -- often stored as attribute-value pairs. Each pair of IEs is linked by a similarity arc representing how close the entities are to be thought. Each IE again points via some relevance arc to the case it constitutes. Thus, cases are represented by a collection of features being more or less relevant and the similarity between the features is represented independently of their relationship in a particular case. NEWLINENEWLINENEWLINEThis basic idea has many appealing properties: For example, very different similarity measures, eg. weighted sums as a typical example, can be calculated on the case-level based on the `elementary' similarities represented at the level of the information entities. The associative structure of the memory facilitates an incremental update and modification of a case memory, new features can be added without reorganizing the entire case base. Finally, case retrieval nets seem to offer a great potential for a parallelized and distributed implementation of case-based reasoners. The limitations of similarity-based retrieval are also discussed -- the most similar case might not be the one that is the easiest to adapt to the needs of a user. Therefore, the design and engineering of case-based systems is still a sophisticated task. Several extensions to case-retrieval nets get also discussed. First, the explicit representation of similarity relationships between each pair of information entities is a quite costly undertaking and can be overcome by defining similarity over a conceptual taxonomy. Second, the case representation can be augmented by so-called microfeatures and also linked to an object-oriented representation. NEWLINENEWLINENEWLINEThe main purpose of case-retrieval nets is to support the fast retrieval of similar cases -- on the theoretical side, retrieval in these nets is worst-case linear in the size of the knowledge base. A spreading activation like inference mechanism is introduced, which leads to very satisfying results and which can be further speed up by several heuristic techniques that significantly reduce the number of cases that have to be considered during retrieval. Error bounds defined for the various heuristics help in judging the quality of the retrieval results. Finally, the problem of how to create a large case base is addressed. NEWLINENEWLINENEWLINEHere, the author presents a very pragmatic and convincing solution. Case bases are automatically generated from existing relational data bases. The task of the knowledge engineer is to define a mapping from the relational data to the case format. Quite often, a case will correspond to a record in a data base table. The similarity arcs between the IEs are computed based on a similarity model, the definition of which nevertheless might require some major effort. Similary, defining the relevance of an IE for a particular case is a laborious task. Two detailed case studies from industrial development projects using case retrieval nets conclude the book. They give an interesting insight into the practical usability of the method and provide rich experience in engineering case-based systems. A general overview over case-based reasoning techniques complements this interesting new method.