PRIVACY PRESERVING ASSOCIATION RULE MINING ON VERTICALLY PARTITIONED DATABASE

Abstract

Association rule mining is a data mining technique used to find interesting associations among . a large set of data _items. For finding association rules from market-basket databases customer-buying habits between the different items (that customers place in their shopping basket) are analyzed. The discovery of such associations can help retailers develop market strategies by gaining insight into which items are frequently purchased together by customers. -Sometimes these association rule mining results disclose some new implicit information about individuals which is against privacy policies. In . vertically distributed databases, the data is vertically partitioned among various sites. These sites wish to work together- to find globally valid association rules without revealing individual transaction data. So some privacy-preserving method must be used, which protect the privacy of the distributed databases and at the same time gives accurate association rules. In this thesis, we propose an algorithm for finding association rules from vertically distributed Boolean databases which maintains a balance between the accuracy of the mining results and the privacy of the databases. For. preserving the privacy, database • is distorted by XORing the boolean data with a. boolean random variable, and then adding some fake transactions in the distorted database. All frequent itemsets are generated for Master's partition. Then intersection of the TIDs of frequent itemsets of Master and real TIDs of other partitions is done. If the intersection value is greater than or equal to some minimum support value (provided by Master Partition) only then the algorithm proceeds. Then the partitions are combined only for the TIDs of Master's partition. Then association rule mining is done by on the combined database and a set of the relative TIDs are made for each candidate itemset. Then again the intersection is performed by third party for each set of TIDs of frequent itemsets to check whether the. itemset is frequent in the real TIDs or not. If the third party sends `OK' then association rules are generated from the frequent. itemsets.