Detecting Class-Independent Linear Relationships within an Arbitrary Set of Features

Classifiers for surveillance sonar systems are often designed to operate on large sets of predefined clues, or features. Sometimes the mathematical definitions for these features are poorly known. Other times the designer is not aware that a fixed and class-independent linear (or affine) relationship exists between subsets of features. We discuss a method based on Gram-Schmidt orthogonalization which allows the classifier designer to determine whether subsets of features have such relationships. Certain features can then be shown unnecessary by application of Wozencraft and Jacobs\´ "Theorem of Irrelevance". An approach is also described to rank features to aid in the selection of an effective subset.