Association Rules Mining: Application to Large-Scale Satisfiability

Big data mining is a promising technology for industrial applications such as business intelligence and marketing. In this paper, we attempt to transfer this background to a scientific domain, which is problem solving. An enormous research investment has been done over the past years for solving NP-complete problems. The exact methods yield the optimal solution if it exists but they are not able to tackle problems of large instances because they would generate a combinatorial explosion and a timeout calculation. Approximate approaches are mostly stochastic and do not guarantee to find a solution even if one exists. For both alternatives, and in order to reach more closely any optimal solution, the search space should be explored in a judicious manner. If we consider the solutions collection as a data set, data mining techniques may help acquiring some knowledge on the landscape fitness function and then exploiting the achieved information in solving the problem. We are especially interested in locating those regions that are promising in the sense they contain solutions of good quality. We aim especially at developing a two-phase approach to address the satisfiability problem in order to cope with its complexity. In the first step, a preprocessing of clauses is undertaken using association rules mining. The achieved associative clauses classes are then exploited to solve the instance. Extensive experiments are performed on BMC benchmarks and numerical results show the benefit of our proposal.