A Novel Error-Tolerant Frequent Itemset Model for Binary and Real-Valued Data

Abstract

Frequent pattern mining has been successfully applied to a broad range of applications, however, it has two major drawbacks, which limits its applicability to several domains. First, as the traditional 'exact' model of frequent pattern mining uses a strict definition of support, it limits the recovery of frequent itemset patterns in real-life data sets where the patterns may be fragmented due to random noise/errors. Second, as traditional frequent pattern mining algorithms works with only binary or boolean attributes, it requires transformation of real-valued attributes to binary attributes, which often results in loss of information. As many of the real-life data sets are both noisy and real-valued in nature, past approaches have tried to independently address these issues and there is no systematic approach that addresses both of these issues together. In this paper, we propose a novel Error-Tolerant Frequent Itemset (ETFI) model for binary as well as real-valued data. We also propose a bottom-up pattern mining algorithm to sequentially discover all ETFIs from both types of data sets. To illustrate the efficacy of our proposed ETFI approach, we use two real-valued S.Cerevisiae microarray gene-expression data sets and evaluate the patterns obtained in terms of their functional coherence as evaluated using the GO-based functional enrichment analysis. Our results clearly demonstrate the importance of directly accounting for errors/noise in the data. Finally, the statistical significance of the discovered ETFIs as estimated by using two randomization tests, reveal that discovered ETFIs are indeed biologically meaningful and are neither obtained by random chance nor capture random structure in the data.