Abstract:

A variety of differentially private FIM algorithms have been proposed. However, current solutions for this problem cannot well balance privacy and data utility over large-scale data. This paper proposes a new differentially private FIM algorithm(TrunSuper). This algorithm truncates the transaction datasets to reduce the dimension, and sorts the items in decreasing order, then eliminates the items with less support. In this way, it can reduce the information loss of the frequent itemsets. This paper also theoretically proves that TrunSuper can produce reasonably accurate results while satisfying differential privacy. Experiments on several real datasets shows that TrunSuper performs better than other previous solutions.

Key words: frequent itemsets mining, differential privacy, transaction truncating, Laplace mechanism