Discovering Overlapping Quantitative Associations by Density-Based Mining of Relevant Attributes

Association rule mining is an often used method to find relationships in the data and has been extensively studied in the literature. Unfortunately, most of these methods do not work well for numerical attributes. State-of-the-art quantitative association rule mining algorithms follow a common routine: (1) discretize the data and (2) mine for association rules. Unfortunately, this two-step approach can be rather inaccurate as discretization partitions the data space. This misses rules that are present in overlapping intervals. In this paper, we explore the data for quantitative association rules hidden in overlapping regions of numeric data. Our method works without the need for a discretization step, and thus, prevents information loss in partitioning numeric attributes prior to the mining step. It exploits a statistical test for selecting relevant attributes, detects relationships of dense intervals in these attributes, and finally combines them into quantitative association rules. We evaluate our method on synthetic and real data to show its efficiency and quality improvement compared to state-of-the-art methods.