Constraint-Based Pattern Set Mining

Abstract

Local pattern mining algorithms generate sets of patterns, which are typically not directly useful and have to be further processed before actual application or interpretation. Rather than investigating each pattern individually at the local level, we propose to mine for global models directly. A global model is essentially a pattern set that is interpreted as a disjunction of these patterns. It becomes possible to specify constraints at the level of the pattern sets of interest. This idea leads to the development of a constraint-based mining and inductive querying approach for global pattern mining. We introduce various natural types of constraints, discuss their properties, and show how they can be used for pattern set mining. A key contribution is that we show how well-known properties from local pattern mining, such as monotonicity and anti-monotonicity, can be adapted for use in pattern set mining. This, in turn, then allows us to adapt existing algorithms for item-set mining to pattern set mining. Two algorithms are presented, one level-wise algorithm that mines for all pattern sets that satisfy a conjunction of a monotonic and an anti-monotonic constraint, and an algorithm that adds the capability of asking topk queries, We also report on a case study regarding classification rule selection using this new technique.