A Generic and Declarative Method for Symmetry Breaking in Itemset Mining

Abstract

The search of frequent patterns in a transaction database is a well-known problem in the field of data mining. Several specific methods for solving this problem and its variants have been developed in the data mining community. A generic and declarative alternative approach to these targeted and specific methods was recently introduced. It consists in representing the data mining problems as constraint networks, then use an appropriate solver as a black box to solve the encoded problem. For instance, several works expressed the frequent itemset mining problem and its variants as constraint networks or Boolean satisfiability, and thus offer the possibility of using the associated efficient solvers to solve the problem. On the other hand, the symmetry notion was very invested and shown to be efficient in the fields of constraint programming and propositional satisfiability. The principle of symmetry could be exported to other areas where some structures can be exploited effectively. Especially, in the field of data mining where several tasks can be expressed as constraint networks. In this work, we propose a generic and declarative method to eliminate symmetries in data mining problems expressed as Boolean constraints. We show how the symmetries between items of a transaction database can be detected and eliminated by adding symmetries breaking predicates (SBP) to the Boolean encoding of the considered data mining problem.