Parallelization Framework for Exact Hiding

Abstract

In this chapter, we elaborate on a novel framework, introduced in [25], that is suitable for decomposition and parallelization of the exact hiding algorithms that were covered in Chapters 14, 15 and 16. The framework operates in three phases to decompose CSPs that are produced by the exact hiding algorithms, into a set of smaller CSPs that can be solved in parallel. In the first phase, the original CSP is structurally decomposed into a set of independent CSPs, each of which is assigned to a different processor. In the second phase, for each independent CSP a decision is made on whether it should be further decomposed into a set of dependent CSPs, through a function that questions the gain of any further decomposition. In the third and last step, the solutions of the various CSPs, produced as part of the decomposition process, are appropriately combined to provide the solution of the original CSP (i.e. the one prior to the decomposition). The generality of the framework allows it to efficiently handle any CSP that consists of linear constraints involving binary variables and whose objective is to maximize (or minimize) the summation of these variables. Together with existing approaches for the parallel mining of association rules [6, 34, 80], the framework of [25] can be applied to parallelize the most time consuming steps of the exact hiding algorithms.