DISTRIBUTED ASSOCIATION RULES MINING USING NONDERIVABLE FREQUENT PATTERNS

Mining association rules in distributed databases is an interesting problem in thecontext of parallel and distributed data mining. A number of approaches have, so far, beenproposed for distributed mining of association rules. However, most of them consider all types offrequent itemsets the same, even though there are different types of itemsets in distributeddatabases, e.g., derivable and non-derivable. In this study, a new application of deduction rules isintroduced for distributed mining of association rules which exploits the derivability of itemsets toreduce communication overhead and to enhance response time. A new algorithm is proposedwhich mines derivable and non-derivable frequent itemsets in a distributed database. Since thecollection of derivable and non-derivable frequent itemsets form all frequent itemsets, ouralgorithm mines all frequent itemsets rather than a subset of them. In the algorithm, there is noneed to scan local databases and exchange messages in order to obtain support counts of derivablefrequent itemsets, since each site can produce them autonomously. Experimental evaluations onhorizontally partitioned real-life datasets show that such exploitation drastically reducescommunication and also improves response time. Therefore the new algorithm is useful whencommunication bandwidth is the main bottleneck.