Machine Learning from Structured Objects

Abstract Machine learning techniques applied to structured objects frequently use a predicate calculus representation to model the world. Unless careful attention is given to the semantics of this model, the results of inductive inference over descriptions of structured objects have unanticipated interpretations. In this paper, a motivation is given for the importance of careful attention to the semantics that underly descriptions of structured examples and categories of such examples. Particular attention is given to the use of must not clauses and the ability to determine relevant attributes. An example from INDUCE/NE is used to illustrate must not clauses with the INDUCE algorithm. An example from CLUSTER/CA is used to illustrate the use of knowledge about relevant attributes in learning.