Mining Disjunctive Rules in Dynamic Graphs

Abstract

Recently, a generalization of association rules that hold in n-ary Boolean tensors has been proposed. Moreover, preliminary results concerning their application to dynamic relational graph analysis have been obtained. We build upon such a formalization to design more expressive local patterns in this special case of dynamic graph where the set of vertices remains unchanged though edges that connect them may appear or disappear at the different timestamps. To design the pattern domain of the so-called disjunctive rules, we have to design (a) the pattern language, (b) interestingness measures which serve as the counterpart of the popular support and confidence measures in standard association rules, and (c) an efficient algorithm that may compute every rule that satisfies some primitive constraints like minimal frequencies or minimal confidences. The approach is tested on real datasets and we discuss the expressivity and the relevancy of some computed disjunctive rules.