The Research On Incremental Learning Algorithm Based On Neighborhood Rough Sets And Applicaiton In Customer Classification

With the rapid development of science and technology, the ability of people to get the data is increasing in recent years. All of sectors of society generate a lot of data, how to extract useful information from these updating data and analysis or process the data has become a difficult problem. The traditional machine learning algorithm is a static learning mechanism, where, all the data must be ready before studying. This requirement in real life cannot be satisfied. However, the incremental learning provides an effective method for solving this problem. So the research on incremental learning mechanism and algorithm for processing data dynamically are of very important practical significance.Neighborhood classified advanced by Hu Qinghua Professor is a simple and effective method, where it is no-incremental classifier and cannot deal with incremental data, which limits its application in some fields. This dissertation found two weaknesses by analysing neighbor classifier:on the one hand, attribute reduction is an important link in the process of classification. The traditional incremental attribute reduction algorithms focus on increasing attributes or increasing samples the without considering these two cases; On the other hand, the performance of classifier depends on information hidden in the training sample. It is hoped that the training sample is as many as possible in the process of classifier learning. But the huge training samples also bring inconvenience:declining the efficiency of classifier. So the innovatory works are as follow.(1) The conception of relative positive region of neighborhood system is proposed. This concept can describe the relation between the thickness of different neighborhood systems.In classification, the thinner a neighborhood system based on a condition attribute induct on universe is, the littler the granularity is, then the more accurate the precision of describing the classification is. In other words, the thinner, the greater the assortment ability of this attribute(2) An incremental attribute reduction algorithm based on improved relative positive region is designed. This algorithm can handle both situations of incremental attributes and samples. When the attribute is added, the conception of relative positive region of neighborhood system can update the original reduction set. When the sample is added, the case contradiction between the additive sample and original reduction set is discussed. In this case, the problem is transformed into the situation of increasing attribute by finding the attributes that can distinguish the new sample and contradiction samples from the complement of original attribute set. So this mechanism can ensure the new attribute reduction set without redundant attributes, and achieve the purpose of combining these two cases.(3) For the problem of low effective of neighborhood classifier, an sample reduction algorithm based on neighborhood rough set is presented, which can be described as:defining the neighborhood radius as the threshold of sample similarity, and taking each training sample as the center, find its largest neighborhood region which covers the largest number of samples with the same category. Then save the center sample and delete the other samples. So that we can not only achieve the result of reduction samples but also improve the classification efficiency.(4) The application of the incremental learning algorithm based on neighborhood rough set in customer classification. Customer classification can achieve accurate classification to the customers and predict the customers’ consumption tendency by using all kinds of technical methods to analysis and mine customer information. This dissertation use the incremental learning algorithm based on neighborhood rough set to achieve the results of accurate classification in the incremental case. The results of experiments show that this algorithm can improve the efficiency of customer classification, and this algorithm has some practical significance.