Data mining comprises two subdisciplines. One of these is based on statistical modelling, though the large data sets associated with data mining lead to new problems for traditional modelling methodology. The other, which we term pattern detection, is a new science. Pattern detection is concerned with defining and detecting local anomalies within large data sets, and tools and methods have been developed in parallel by several applications communities, typically with no awareness of developments elsewhere. Most of the work to date has focussed on the development of practical methodology, with little attention being paid to the development of an underlying theoretical base to parallel the theoretical base developed over the last century to underpin modelling approaches. W e suggest that the time is now right for the development of a theoretical base, so that important common aspects of the work can be identified, so that key directions for future research can be characterised, and so that the various different application domains can benefit from the work in other areas. We attempt describe a unified approach to the subject, and also attempt to provide theoretical base on which future developments can stand.