Mining Time-Profiled Associations: A Preliminary Study

Abstract

A time-profiled association is an association pattern consistent with a query sequence along time, e.g., identifying interacting relationship of droughts and wild fires in Australia with the El Nino phenomenon in the past 50 years. Association patterns by traditional association rule mining approaches reveal the generic dependency among variables, however, the evolution of these patterns along time is not captured. Hence the time-profiled association mining is used to incorporate the temporal evolution of association patterns and identify the co-occurred patterns consistent along time. Mining time-profiled associations is computational challenging due to large size of itemset space and long time points in practice. A naive approach of mining time-profiled associations can be characterized using a two-phase paradigm. The first phase generates the statistical parameter (e.g., support) sequences along time, and the second phase retrieves similar sequences with the query sequence. However, exponentially increasing computational costs of generating all combinatorial candidate itemsets become prohibitively expensive for the previous work. In this paper, we propose a novel one-step algorithm to unify the generation of the sequence of statistical parameters and sequence retrieval. The proposed algorithm substantially reduces the itemset search space by pruning candidate itemsets based on the monotone property of lower bounding measure of sequence of statistical parameters. Experimental results show that our algorithm outperforms the naive approach.