Efficiently Depth-First Minimal Pattern Mining

Arnaud Soulet (1, 2) , François Rioult (3)

1 BDTLN - Bases de données et traitement des langues naturelles
2 LIFAT - Laboratoire d'Informatique Fondamentale et Appliquée de Tours
3 Equipe CODAG - Laboratoire GREYC - UMR6072

Condensed representations have been studied extensively for 15 years. In particular, the maximal patterns of the equivalence classes have received much attention with very general proposals. In contrast, the minimal patterns remained in the shadows in particular because of their difficult extraction. In this paper, we present a generic framework for minimal patterns mining by introducing the concept of minimizable set system. This framework addresses various languages such as itemsets or strings, and at the same time, different metrics such as frequency. For instance, the free and the essential patterns are naturally handled by our approach, just as the minimal strings. Then, for any minimizable set system, we introduce a fast minimality check that is easy to incorporate in a depth-first search algorithm for mining the minimal patterns. We demonstrate that it is polynomial-delay and polynomial-space. Experiments on traditional benchmarks complete our study.

Domaines: Base de données [cs.DB]; Machine Learning [stat.ML]

Arnaud Soulet, François Rioult. Efficiently Depth-First Minimal Pattern Mining. Advances in Knowledge Discovery and Data Mining, 18th Pacific-Asia Conference, PAKDD 2014, 2014, Tainan, Taiwan. pp 28-39, ⟨10.1007/978-3-319-06608-0_3⟩. ⟨hal-01021412⟩