Biclustering with Background Knowledge using Formal Concept Analysis

Biclustering methods have proven to be critical tools in the exploratory analysis of high-dimensional data including information networks, microarray experiments, and bag of words data. However, most biclustering methods fail to answer specific questions of interest and do not incorporate background knowledge and expertise from the user. To this end, query-based biclustering algorithms have been recently developed in the context of microarray data; these algorithms utilize a set of seed genes provided by the user to prune the search space and guide the biclustering algorithm. In this paper, a novel Query-Based Bi-Clustering algorithm, QBBC, is proposed via a new formulation that combines the advantages of low-variance biclustering techniques and Formal Concept Analysis. We prove that statistical dispersion measures that are order-preserving induce an ordering on the set of biclusters in the data. In turn, this ordering is exploited to form query-based biclusters in an efficient manner. Our novel approach provides a mechanism to generalize query-based biclustering to sparse highdimensional data such as information networks and bag of words. Moreover, the proposed framework performs a local approach to query-based biclustering as opposed to the global approaches that previous algorithms have employed. Experimental results indicate that this local approach often produces higher quality and precise biclusters compared to the state-of-the-art query-based methods. In addition, a performance evaluation illustrates the efficiency and scalability of QBBC compared to full biclustering approaches and other existing query-based approaches.