Variable-Scale Clustering For Decision Making

Facing the problems of developing decision analysis systems under the cross-industry standard process for data mining(CRISP-DM),that is the data structure determination,analytic hierarchy transformation and analytic results verification,this paper focuses on the basic data scale determination,the variable-scale data analysis mechnasim during the decision making process,as well as the data mining application techniques based on the variable-scale data analysis.The main pratical problems in developing decision analysis systems are as follows.(1)The data structure determination problem for decision making.According to the CRISP-DM,the primary task in developing decision analysis systems is to determine the data anlysis subjects following the business analysis demands.That how to establish the analysis subjects-supported data structure has great influence on the efficiency of data mining algorithms and the quality of results.(2)The analytic hierarchy transformation problem for decision making.Since the pratical business itself has multi-scale characteristics,in order to reach the goal of reasonable decision-making,managers need to keep transforming the analytic hierarchies in decision making,which is essentially the transformation of data analytic hierarchies.Hence,how to automatically recognize the reasonable business data analysis hierarchiey is the key to reducing the complexity of the decision-making process as well as improving the quality of decision-making results.(3)The analytic results verification problem for decision making.The analytic results verification that aims to ensure the consistency of decision analysis results at different hierarchies,is the measurement standard of the end of a decision analysis process.After clarifying the pratical problems above,the research problems studied in this paper include:(1)The basic data scale determination.That studies the data representation method of multiple hierarchies for decision analysis,as well as how to determine the scale structure of basic business data,following the characteristics of three original business data types,that is categorical data,binary data and numerical data,in order to establish a complete data structure for variable-scale data analysis.(2)Variable scale data analysis mechanism in the process of decision analysis.That studies data scale transformtion mechanism based on the data mining results through simulating managers' thinking processes during the transformation of the decision analysis hierarchies,in order to improve the quality of decision results.The influence caused by different data types of original business data to the scale transformation is also be considered,which achieves the automatic recognition and transformation of reasonable business data analysis hierarchies.The automatic decision analysis mechanism based on the data scale transformation is established.(3)Data mining application tecniques based on the variable-scale data analysis.According to the actual management business of Grit Tao Sports Development Co.,LTD,Sina Weibo social media platform and China Academy of Launch Vehicle Technology,the application techniques of the scale transformation mechanism is studied.The contribution of this research work is as follows.(1)Establishing a multi-scale data model to represent the problem solving space of decision problems,which could describe the basic business data scale composition and relation of all candidate analysis hierarchies of decision problems,and provide a complete data structure for the transformation of decision analysis hierarchies.The existing data structure model takes the initial data scale of business data acquisition as the basic data scale for data mining,which makes the single scale data model unable to support the hierarchical transformation requirements of decision analysis.(2)The data scale transformation strategy and scale transformation mechanism for the rational decision-making thinking process are proposed,which could determine the reasonable data analysis hierarchy and data scale transformation path based on decision analysis results.When there is a mismatch between the data mining algorithm results and the decision analysis hierarchy,the existing data mining method is just fully dependent on analysts to transform the business data analysis hierarchy subjectively.This research work build the automated data scale transformation mechanism of the CRISP-DM core data mining process,that data preparation,modeling and evaluation,which could adjust the business data analysis hierarchy of the data preparation and modeling process based on the evaluation results,and optimize the scale transformation strategy through the quantitative scale transformation value of different attributes in the multi-scale business data model.(3)The variable-scale clustering method(VSC)is proposed,whose accuracy is guaranteed by the scale transformation theorem and the satisfaction class consistency theorem.The comparative experiments show that the clustering results of the variable-scale clustering method are effective and insensitive to the initial algorithm parameter.According to the proposed variable-scale clustering method,this paper carries out relevant application research on three practical management business scenarios including:?As for the decision analysis problem with duplicate categorical data,a variable-scale clustering algorithm for duplicate categorical data is proposed.The experimental results using the marathon data from Grit Tao Sports Development Co.,LTD show that the proposed method could support making differentiated management scheme for runners.?As for the decision analysis problem with both categorical data and binary data,a hybrid variable-scale clustering algorithm for categorical and binary data is proposed.The experimental results using the customer data from Sina Weibo social network platform show that the proposed method could support making differentiated management scheme for customers.?As for the decision analysis problem with numerical data,a time-based variable-scale clustering algorithm for numerical data is proposed.The experimental results using the material inventory data of space products from China Academy of Launch Vehicle Technology show that the proposed method could support making differentiated management scheme for materials.