Mining the Preference Relations and Preference Graphs

Abstract

When knowledge mining starts from contingency tables, many types of knowledge become apparent that would otherwise went unnoticed. In this paper we start from contingency tables for pairs of attributes whose domains are ordered. The domain of each attribute can be interpreted as a preference list. We present a common 2-d pattern that can be collectively called preference relation. Preference relation tells that one of the choices is accepted to a higher degree over another. On one end of the spectrum, a weak preference borders equivalence relation between choices. On the other, a very strong preference is similar to subset relation. We present several tests that can distinguish various forms of preference relation knowledge and also subset and equivalence. Our experience in data mining with the application of the 49er system shows that the exploration of many databases frequently leads to large numbers of preference-type regularities. Large numbers of preference-type regularities can be combined into concise, useful forms of preference graphs. We compare preference graphs to taxonomies and inclusion graphs. We illustrate the presented algorithms by applications on the International Social Survey Program (ISSP) databases.