Probabilistic classification in Osiris, a view-based OO DBMS and KBMS

The Osiris system implements an object data model which was designed to enable the users to share data through a view mechanism. Views, which play the role of subclasses in an object perspective, are defined by their logical properties. Determining the views whose properties are satisfied by a given instance is the very mechanism of classification in knowledge bases. An in-depth static analysis of the views and the assertions defining their logical properties allows the construction at compile time of a network to classify the objects of the base. Contrary to most neural networks used for classification, the architecture of this network (arcs, nodes, and weights) is fully determined by the compilation process, and it needs no learning. This network can also achieve the classification of incompletely known objects and calculate the probability of each potential view, provided that probabilistic information is given on the domains of the missing attributes. The relationship between the complexity of the network and the kind of assertions which is taken into account is studied