Using Multidimensional Projection to Find Relations

Abstract

When shortage of domain knowledge prevents us from chosing good attributes to represent the examples, learning is difficult. Expressing the target concept using only primitive (low-level) attributes may be complex, and the individual contribution of each attribute to the target's definition becomes insignificant. This aggravates attribute interaction (a situation in which complex relationships among attributes appear in the target concept). Then the learner needs to find relations and use them to help learning. This paper purports the relational operator projection as a useful tool to find relations. We describe MRP, a learning algorithm based on multidimensional relational projection, and compare it empirically to other learning systems. In spite of its simple search strategy, MRP performs well on synthetic concepts and real-world data. This advantage is attributed to the projection operator achieving the required functionality: finding relations.