EÆcient Discovery of Concise Association Rules from Large Databases

Association rules are interesting correlations among attributes in a database. These rules have many applications in areas ranging from e-commerce to sports to census analysis to medical diagnosis. The discovery of association rules is an extremely computationally expensive task and it is therefore imperative to have fast scalable algorithms for mining these rules. In this thesis, we present eÆcient techniques for discovering association rules from large databases and for removing redundancy from these rules so as to improve the quality of output. We also handle growing databases. Speci cally, we present three new algorithms: (1) ARMOR: This algorithm discovers association rules from databases and requires at most two database scans. We empirically show its performance to be within a factor of two of an unachievable lower bound. (2) g-ARMOR: This is an extension to ARMOR that is designed to remove redundancy from association rules during the mining process. This is especially important because the number of association rules generated in typical mining operations runs into the tens of thousands. g-ARMOR results in an orders of magnitude reduction in the number of rules thereby making the mining output comprehensible to end users. (3) DELTA: This algorithm incrementally mines evolving databases. It utilizes previous mining results to eÆciently mine the current database after it has been updated with fresh data. It also handles situations where the mining speci cations over the current database di er from those used over the original database, a common occurrence in practice.