Using typicality to support inference and learning

This paper presents a methodology to support inference and learning using typicality. The methodology views typicality as a trait of specific concept characteristics and of concept themselves. We build up our effort upon earlier work by Collins and Michalski [10], Kahneman and Tversky [22], Vignes and Lebbe [48] and others. The paper proceeds by first describing the evidence which led to the necessity for typicality modeling and the relationship of typicality metrics with probabilistic measures. It then conceptualizes typicality and develops models to support its computational implementation. The backbone of our work lies in the assessment of weight values of attributes that are used to represent concepts in intension. We use these values to support learning of concept identification rules, to assess similarity between concepts and to draw inferences about value patterns. We demonstrate our approach by way of a ‘real world’ case study involving anemia type identification. To implement this case study we have developed a typicality based system, namely TYPOS. The final section discusses the pros and cons of the proposed methodology and ways in which this work can be further extended.