Compressing Closed Frequent Itemsets with Controlled Information Loss

Abstract

Closed frequent itemsets (CFIs) condense frequent itemsets without loss of information. For large and dense datasets like big data and unbound big data streams, even the number of CFIs generated can be enormous. In such scenarios approximation is preferred against an accurate solution. Subset Significance Threshold (SST) is an effective constraint variable in mining significant CFIs. The support of the insignificant CFIs is approximated to the support of their immediate superset. However, few insignificant CFIs are approximated beyond specified SST due to chaining effect. To overcome this limitation in SST, the authors are proposing an enhancement to the SST (e-SST) in this paper to improve the degree of accuracy of the approximated insignificant CFIs. The merging of insignificant CFIs to thier superset is limited to one level so that the approximation is bound within specified SST. Experimental results show that the e-SST technique is efficient than SST in limiting the approximation of the support of insignificant CFIs within the specified threshold, thus reducing the information loss.