Research On Spatio-temporal Association Rule Mining

Spatio-temporal association rule mining (STARM) is a frontier issue in spatial data mining. With the development of geoinformation technology and the accumulation of spatio-temporal data, STARM has become an important tool for spatio-temporal data analysis and mining. Theory, methods and tools for STARM play an important role in both research and practice. The following issues are discussed in this study:This paper discusses the basic concepts of STARM from the point of view of probability and first order logic, including items, itemsets, association rules, support and confidence.Based on probability theory and the fact that the support of an itemset is the probability of the corresponding event, the formal definition of item, itemset and data is proposed. According to Axiom1(a dataset is ready for association rule mining iff every itemset’s support is calculable), Theorem1, which can be used to judge the feasibility of the mining task over a certain dataset, is proposed. An association rule is defined as an expression of conditional probability. The size of a dataset and the support count of an itemset are defined using a finite measure, and it can be proved that the support count of an itemset divided by the size of data equals to its support. The apriori proerty, which is widely used in association rule mining algorithms, is proved as Theorem2. Also, the definition of spatio-temporal association rule is proposed. These axioms, definitions, theorems and corollary form the foundation of STARM and rule evaluation methods.This paper proposed a method for association rule mining which supports a variety of data types. The basis of this method is the General Association Rule Mining Framework (GARMF), the main issues include the calculation of data size and support count in transaction, spatial and spatio-temporal data types. An apriori-like algorithm is integrated into this method to assure its efficiency; a fast intersect method for spatial datasets and strategies for incremental maintenance of discussed rules are also proposed for this method. The experimental results with land cover data show that our method is available to mine spatio-temporal association rules from transaction, spatial and spatio-temporal data, and can be used to extract and analyze land cover change trajectories. Not all of the results obtained from STARM are correct, interesting and applicable, it is necessary to evaluate and filter them. Common measures of itemsets and association rules are introduced, and a similarity measure is proposed. Similarity is a subjective measure. It takes relationships between items into consideration when comparing the mining results with the knowledge base. The filtering methods based on item-constraints and measures are also discussed, and Rule Filtering Library (RFL) is implemented to simplify the measuring and filtering procedure.A STARM support system named DAP Shell is implemented with GARMF and RFL. This system can be used to mine spatio-temporal association rules from transaction and spatial data. It can also be used to evaluate and filter the frequent itemsets and association rules.