Abstract: Multi-scale mining has been applied in the fields of graphic images,geographic information,signal analysis,data mining,etc,and also has related research and application in the fields of association rules,clustering and classification mining.Nevertheless how to divide datasets into common scales and how to construct multi-scale datasets have not been studied in depth.Starting with the task of multi-scale data mining,this paper defined the concept of scale and gave a multi-scale dataset model and a benchmark scale scoring model.This paper proposed a multi-scale partition algorithm based on the discretization method of probability density estimation,which extends the data types of divisible scales,and its partition results are closer to the multi-scale characteristics of data with lower time complexity.This paper also proposed a multi-scale dataset method,a multi-scale data set algorithm and a benchmark scale selection algorithm.Multi-scale entropy and information entropy were used as evaluation methods.On the basis of expanding the multi-scale dataset method,the scale effect produced by the meso-scale derivation of multi-scale data mining can be effectively reduced,and the time complexity can be controlled.The proposed algorithm and model were validated and analyzed by using the real population dataset of H province,UCI common dataset and IBM dataset.The experimental results show that the proposed method is feasible and the proposed model is effective.The application of the proposed methods improvescoverage by 1.6%,F1-measure by 2.1% andaccuracy by 3.7% in scale deduction process,and has low average support error.

Key words: Construction of multi-scale datasets, Discretization, Information entropy, Multi-scale data mining, Multi-scale entropy, Multi-scale scaling, Reference scale selection