Learning Context Dependent Concepts

This article presents the concept of context dependent classiication. In context dependent classiication of elementary objects, an elementary object is represented by a node in a graph. A node can be identiied to belong to a certain node class, if one of the characterizing contexts for that class can be embedded in the graph properly. Thus, the classiication of a node does not only depend on properties of the elementary object itself, but also on properties of its environment. For learning a context dependent classiier, the characterizing contexts have to be learned in addition to the relevant properties of the node itself. A context will be approximated by paths of a certain length. This way the context can be described by feature vectors and well-known decision tree induction can be used. For induction, the context depth is increased until the classes can be separated suuciently. In this article, the context dependent classiication of elementary objects is generalized to a context dependent classiication of tuples of elementary objects and of substructures. The classiication of structured objects is a special case of the context dependent classiication of substructures. Learning context dependent classiications can be important in many applications.