IP Cell Suppression Model Reduction and a Top Down Design

Most Economic Census (EC) products use cell suppression to avoid disclosure in publication. We have a sophisticated sequential linear programming (Seq-LP) cell suppression program that is as efficient and robust as that model can provide. The sequential approach is a process restricted by the linear programming (LP) model. Therefore, it doesn’t provide a global optimal solution and sometimes oversuppresses. The ideal model is a simultaneous integer programing (IP) algorithm. The simultaneous IP (Sim-IP) provides an ideal solution without oversuppression.
Sim-IP cell suppression is well known for providing a global optimal solution. However, the complexity of the model makes it impractical to use in a production environment. The complexity of the model, that is largely determined by the number of primaries (targets) and solution space (available cells to complement primaries), grows exponentially as the amount of data being processed increases. In other words, Sim-IP scales poorly, and can only process small datasets. Over the years, researchers have been trying to solve the cell suppression problem with IP, but only found it to have a limited application.
Due to the characteristics of Economic Census data, the cell suppression model is usually sparse. In addition, the data are widely duplicated and there are many small, sensitive cells with small protection requirements. Our first step in this research is to do a limit test: explore what data size the Sim-IP can endure in terms of targets and solution space. Then, we will offer remedies to reduce the IP model size and achieve solutions for larger tables. With our understanding of Census data characteristics, we can reduce model size. We employ a top-down approach that decomposes the data into a geographic hierarchical space. Within this framework we examine three remedies: 1) restriction of the solution space, 2) restricting the number of targets, and 3) a combination of 1 and 2. With Sim-IP cell suppression, we aim for solving a medium-large Economic Census problem.