ABBA: Adaptive bicluster-based approach to impute missing values in binary matrices

Missing values frequently pose problems in binary matrices analysis since they can hinder downstream analysis of the datasets. Despite the presence of many imputation methods that have been developed to substitute missing values with estimated values, these available techniques have some common disadvantages: they need to fix some parameters (e.g., number of patterns, number of rows to consider) to estimate missing values - with little theoretical support to determine these parameters -; and, missing values need to be recomputed from scratch as parameters change. In this paper we propose a novel algorithm (ABBA: Adaptive Bicluster-Based Approach) that does not have the above limitations. Further, a formal framework that justifies the rationales behind ABBA is detailed. Finally, experimental results over both synthetic and real data confirm the viability of our approach and the quality of the results, that overcomes the ones achieved by the main competing algorithm (KNN). © 2010 ACM.