Case-driven collaborative classification

The Multiple Classification Ripple Down Rules (MCRDR) knowledge acquisition approach is explored in an environment where knowledge is continuously changing and where multiple agents need to contribute to and learn from the resultant expert system. A collaborative approach to expert system development is proposed in which private individual views and public shared views of the knowledge can coexist, permitting changes in the knowledge to be highlighted, the impact notified to interested parties, and conflicts between agents to be exposed and resolved. By tracking the public and private historical case-RuleNode associations, the proposed approach allows consensus to be built; hence the brittleness of the acquired knowledge can be more rapidly reduced. The support centre of a major international corporation in the information and communications industry forms a case study in which the nature of trouble-shooting (problem solving) is studied. The research finds that troubleshooting comprises a case-configuration classification- conclusion cycle. An analogy between collaborative MCRDR and collaborative tagging systems is developed and a stochastic model is derived showing that the trajectories in much of the past machine learning case-driven knowledge acquisition studies can be predicted by the acquisition of classification knowledge on the basis of a random set of incoming repetitive cases, irrespective of the specific case-driven knowledge acquisition approach employed. Further to this, the research highlights the significant amount of domain dependent Knowledge Engineering expertise that can be required by various Ripple Down Rules (RDR) implementations, and proposes that rather than primacy being given to the classification as in conventional rule based systems, or alternatively to the case as in existing RDR systems, primacy should be given to building consensus between collaborating agents in as far as consensus is required to achieve a mature knowledge base. A collaborative hybrid Case And Rule Driven (CARD) approach to knowledge acquisition known as 7Cs is proposed. The 7Cs approach supports the Collaborative Configuration and Classification of a stream of incoming problem Cases via a set of ConditionNodes linked to their Classes and associated Conclusions. The approach is trialled through a software prototype system known as FastFIX. After less than a day of collective effort the test team had acquired enough knowledge for FastFIX to automatically identify and locate solutions to approximately 90% of problems in the selected sub-domain.