Comparação da redução de dimensionalidade de dados usando seleção de atributos e conceito de framework: um experimento no domínio de clientes

Abstract

Information related to the Customers at companies are collected and stored in databases. The administration of these data often requires the use of a computational tool. The building of a Customer Profile model from the database requires the process of knowledge discovery in databases. This search of knowledge and extraction patterns of the databases demands the use of a tool with analytics capability to extract information that are implicit, and are previously unknown, but, potentially useful. A data base through of the recovery of date, obtain information of the Customers, but the difficulty is in the fact of these systems do not generate patterns. However, these databases have an expressive amount of data, where redundant information it prejudices this process of patterns extraction. Thus, dimensionality reduction methods are employed to remove redundant information and improve the performance of the learning processes the speed as in the performance of classifier. Furthermore, it identifies a subset of relevant and ideal attributes for a determinate database. The two methods of dimensionality reduction used in this search were: Attribute Selection and Framework Concepts which theretofore were not applied in Customer domain. The Attribute Selection Method has as goal to identify the relevant attributes for a target task, taking into account the original attributes. Considering the Framework Concepts it promotes successive refinements on the attributes where can tale he building of a model more consistent application domain. The present search applied these two methods in order to comparison of these in the Customer domain, using three databases called: Stalog, Customer e Insurance. This paper identified five main steps in order to comparison of the two methods: Preparation of Database, Choice of Database, Application of the Attributes Selection and Framework Concepts Methods, Execution of the Algorithms of the Classification and Evaluation of the Results. With the implementation of theses five steps composed of several processes, it was possible to compare the two methods and identify the best classifiers algorithms and consequently to create the attributes more relevant for a database, increasingthe performance of the learning process. Of this way, with the best subset identified is possible submit them to the application of the Data Mining Tasks which allow the building of rules that help the Knowledge Management of Customer Profile.