Mining Decision Rules From Multi-source Information System

Existing multi-source data mining usually focused on association rules mining in transactional database,and the mining of decision rules from multi-source decision information system is little reported.In an era of explosive growth of data,the collection and storage of data gradually showing some characteristics,such as multi-source,high-dimensional and heterogeneous.According to these characteristics,the centralized processing way cannot meet some specific requirements,and may face several problems such as inconsistent data format,large amount of data transmission,and disclosure of data privacy.Therefore,the way to generate global rules by mining local rules from different data sources,which provides a new way to solve the knowledge discovery from multiple information sources.Meanwhile,by defining multiple metrics,it is possible to effectively measure high-voting decision rules,exceptional rules,and other meaningful decision rules from multi-source information system.In this paper,a format definition of decision rules based on neighborhood granulation is given,the mining models and algorithms for high-voting decision rules,exceptional rules,and global rules from multi-source information system are designed.Finally,the effectiveness of the proposed algorithm is verified by experiments.This dissertation first introduces the related work of pattern mining from multiple related transaction databases,and analyzes the related technologies.In addition,this dissertation introduces some basic theories such as neighborhood granulation,large margin,rule learning.Then,based on the neighborhood granulation,this dissertation studies the mining of all kinds of decision rules from multi-source decision information system.The main research results are as follows:(1)A formal presentation of decision rule via the sample's neighborhood granulation is constructed.On this basis,various metrics about decision rule such as coverage,number of votes are defined to mine high-voting decision rules that meet these metrics.Experimental results demonstrate the effectiveness of the proposed algorithm.(2)We firstly construct the formal presentation of decision rules via the sample's neighborhood granulation.Then,a variety of metrics such as cover degree,important degree and exceptional deviation degree for the decision rule are defined,which can be used to mine exceptional rules from multi-source information systems.The experimental results demonstrate the mining process and results of the exceptional decision rules.(3)A formal presentation of decision rule via the sample's neighborhood granulation is constructed.Based on this,the weight of each information source is measured based on the consistency between information sources.Finally,a weighting model for mining global decision rules via synthesizing all local decision rules is proposed.Extensive experimental results demonstrate that the proposed decision rules synthesization model is effectiveness and scalable.