Hierarchical Conceptual Clustering in a First Order Representation

This paper presents work in progress on unsupervised learning from observations represented in a subset of first order logic. In most systems, observations are described by pairs (attribute, value) and two main approaches can be distinguished: bottom-up methods that consider all the observations and build a hierarchy of concepts starting from the observations and generalizing them/top-down incremental methods that process observations one after the other and incorporate them by a search down the current hierarchy. Our aim is to test how incremental methods could be adapted to representations written in first order logic and we intend to test several similarity measures and generalization methods in order to study their appropriateness.