The Multi-resolution Co-location Miner: A New Algorithm to Find Co-location Patterns in Spatial Dataset

Given a olle tion of boolean spatial features, the o-lo ation pattern dis overy pro ess nds the subsets of features frequently lo ated together. For example, the analysis of an e ology dataset may reveal the frequent o-lo ation of a re ignition sour e feature with a needle vegetation type feature and a drought feature. The spatial o-lo ation rule problem is di erent from the asso iation rule problem. Even though boolean spatial feature types (also alled spatial events) may orrespond to items in asso iation rules over market-basket datasets, there is no natural notion of transa tions. This reates diÆ ulty in using traditional measures (e.g. support, on den e) and applying asso iation rule mining algorithms whi h use support based pruning. In our re ent work [22℄, we proposed a notion of user-spe i ed neighborhoods in pla e of transa tions to spe ify groups of items in [22℄, new interest measures for spatial o-lo ation patterns whi h are robust in the fa e of potentially in nite overlapping neighborhoods,and an algorithm to mine frequent spatial o-lo ation patterns and analyzed its orre tness, and ompleteness. The Co-lo ation Miner generates andidate prevalent o-lo ations in the spatial feature level and generates table instan es for the andidate o-lo ations to he k their prevalen e. When the false andidate prevalent olo ation set is large, the performan e of the Co-lo ation Miner de reases. Due to spatial auto orrelation, the lo ations of individual spatial features of a point data set are often lustered spatially, the Co-lo ation Miner is omputationally expensive without taking spatial auto orrelation into onsideration. In this paper, a new algorithm alled Multi-resolution Co-lo ation Miner is presented. The proSupported in part by the Army High Performan e Computing Resear h Center under the auspi es of Department of the Army, Army Resear h Laboratory Cooperative agreement number DAAH04-95-2-0003/ ontra t number DAAH04-95C-0008 posed algorithm has two logi al phases, namely lter and re nement. The lter phase summarizes the original point dataset into a smaller latti e dataset using spa e partitioning whi h allows the omputation of the upper bounds of the interest measures. It eliminates many non-interesting o-lo ations, redu ing the set of andidates to be explored by the re nement phase, whi h omputes the true values of the interest measures. We show that the proposed algorithm is orre t and omplete and the proposed algorithm is several times faster than the traditional Co-lo ation Miner algorithm in a dataset with spatially auto orrelation by experiments.