Association Rules in Incomplete Databases

Abstract

Discovering association rules among items in large databases in recognized as an important database mining problem. The problem has been introduced originally for sales transaction database and did not relate to missing data. However, missing data often occur in relational databases, especially in business ones. It is not obvious how to compute association rules from such incomplete databases. It is provided and proved in the paper how to estimate support and confidence of an association rule induced from an incomplete relational database. We also introduce definitions of expected support and confidence of an association rule. The proposed definitions guarantee some required properties of itemsets and association rules. Eventually, we discuss another approach to missing values based on so called valid databases and compare both approaches.