STAR: Summarizing Timed Association Rules

Timed association rules (TARs) generalize classical association rules (ARs) so that we can express temporal dependencies of the form “If X is true at time t , then Y will likely be true at time (t + τ ).” As with ARs, solving the TAR mining problem can generate huge numbers of rules. We show that methods to summarize ARs cannot work directly with TARs, and we develop two notions—strong and weak summaries—to summarize a set of TARs. We show that the problems of finding strong/weak summaries are NP-hard, and we provide polynomial-time approximation algorithms. We show experimentally that the coverage provided by our summarization methods is very high. Both technical measures based on coverage and human experiments on six World Bank datasets using 100 subjects from Mechanical Turk and a separate experiment with terrorism experts on a terrorism dataset show that while both summarization methods perform well, weak summaries are preferred, despite their taking more time to compute than strong summaries.