THE PROBLEM OF ATTRIBUTE‐WEIGHTING IN NUMERICAL CLASSIFICATION

Publisher Website
Google Scholar
Abstract
Summary: Serially dependent attributes arise when a primary attribute controls a string of secondary attributes, so that the system takes the form (X or not‐X; if X, then Y or not‐Y, Z or not‐Z; otherwise question inapplicable). Kendrick (1965) pointed out that in numerical classifications, unless steps are taken to apply differential weighting to primary and secondary attributes, the secondaries may over‐ride the primaries, and that this may result in unacceptable classifications. His suggestions are re‐examined in the light of modern metric and information‐statistic programs; it is concluded that the suggestions are numerically sound, but require modification in the context of modern programs. Essentially, the revised suggestions consist in reducing the contribution of secondary attributes by dividing by the number of secondaries in the string, and then taking the number of secondaries as zero when consolidating into an overall similarity measure.

This publication has 4 references indexed in Scilit: