Association rules mining is an important part of data mining, whose ultimate goal is to seek the potential and frequent patterns or correlations hidden behind data. Multi-scale science is an emerging research field whose essence is as follows: analyzing the multi-level and multi-scale structural characteristics of research objects, exploring the root cause of multi-scale expressions and discussing the deep relationships among different scale expressions. Multi-scale theory has been introduced into spatial data mining, and many elementary researches on multi-scale features of spatial data have been done. This paper introduced multi-scale theory into data mining, and further more, pushed it to more extensive data types. Taking association rules mining as the pointcut, this paper conducted a study of universal multi-scale data mining on theoretical and methodological aspects. Referencing and centring on the studying essence of multi-scale science, we researched multi-scale data theory which bases on related concepts as the principal parts. We put forward the process framework of multi-scale data mining, and proposed scaling-up mining algorithm and scaling-down mining algorithm of multi-scale association rules mining on the basis of the theory and framework mentioned above. The proposed algorithms realized multi-scale mining for association rules, and provided theoretical and methodological support for the multi-scale decision of users.This paper took multi-scale association rules mining as the studying essence, whose main contents are as follows:1. Researches on the multi-scale data mining theory.To overcome the limitation that there is still a lack of universal and integrated theoretical foundation in multi-scale data mining field, we conducted a study of multi-scale data mining theory in three major aspects: multi-scale data, multi-scale data mining and multi-scale data mining process framework. Firstly, we put forward the definitions of data-scale-partition, data-scale and unit-scale dataset on the basis of concept hierarchy. Following those definitions, this paper also brought forward four kinds of relationships between multi-scale datasets: ancestor and descendant datasets, father and son datasets, sibling datasets, upper-layer and lower-layer datasets respectively, after which the concept system of multi-scale data was formed preliminarily. Secondly, we gave the definition of multi-scale data mining, illustrated the scale convert for knowledge as the studying essence of multi-scale data mining, classified the multi-scale data mining algorithm into two aspects: scaling-up mining algorithm and scaling-down mining algorithm on the basis of the generalized classification of scale convert, and confirmed the essence and direction of multi-scale data mining. Lastly, we established a multi-scale data mining process framework in stages, which is used to guide and standardize the process of multi-scale data mining.2. The proposal of scaling-up association rules mining algorithm.To make up for deficiencies that there is still no explicit multi-scale data mining algorithm, aiming at association rules mining and centring on scale convert for knowledge, we proposed an algorithm named SU-ARMA(Scaling-Up Association Rules Mining Algorithm) on the basis of sampling theory and Jaccard similarity coefficient, which realized scaling-up convert for knowledge among multi-scale datasets.3. The proposal of scaling-down association rules mining algorithm.Aiming at association rules mining and centring on scale convert for knowledge as well, we proposed an algorithm named SD-ARMA(Scaling-Down Association Rules Mining Algorithm) on the basis of inverse distance weighing in interpolation method. And SD-ARMA realized scaling-down convert for knowledge among multi-scale datasets. And the confidence interval for error rate of SU-ARMA and SD-ARMA was deduced and proved with the help of statistical principle and machine learning theory as well. Further more, we analyzed the advantages of SU-ARMA and SD-ARMA compared with traditional association rules mining methods, and elucidated their applicable domains.4. Verification experiments on the multi-scale data theory and multi-scale association rules mining algorithm.Algorithms of SU-ARMA and SD-ARMA were applied to IBM T10I4D100 K synthetic dataset and demographic dataset from H province whose multi-scale features are obvious. The experimental results turn out that SU-ARMA and SD-ARMA have better coverage rate and accuracy, lower average support error, and their efficiency is also better than traditional way of applying Apriori directly. Algorithms of SU-ARMA and SD-ARMA are feasible and efficient. |