Classification Based On Influence Functions

Classification is one of the important research topics in data mining, it is widely used in areas such as finance, retail, telecommunications, biological industry, and has attracted growing attention from academia and industry. Classification is a supervised learning method, we can obtain the class label of unknown samples by learning through the training data set model. How to improve the accuracy of classification model is the key of classification. Many traditional classification methods have been proposed, such as k-nearest neighbor, decision tree induction, naive bayes, neural network, support vector machine(SVM),and case-based reasoning. These traditional classification methods have the same basic idea: Given the training sample set D, and learn a classification model M on D, then the class label of instance can be determined by the model. These classification algorithms are based on the training data set learning classifier to classify unknown instances, and they are all affected by the training data set.Therefore, on the basis of the classification of the existing classic algorithms, in this paper,we put forward a new classification algorithm from a new perspective, namely classification based on influence function. It is the key to define an appropriate influence function, combined with the four influence functions defined in this paper: Liner Influence Function, Square Influence Function, Exponential Influence Function and Gravity Influence Function, we propose four classification methods based on influence functions: CBLIF(Classification Based on Liner Influence Function), CBSIF(Classification Based on Square Influence Function),CBEIF(Classification Based on Exponential Influence Function) and CBGIF(Classification Based on Gravity Influence Function). Strictly, classification based on influence function is not a classification method, but a kind of classification paradigm. Define different influence function can lead to different classification algorithms. We can unify the existing classification algorithm to a framework, if we define an appropriate influence function, it can allow arbitrary shaped boundary, and there is no need to learn boundary. It will bring great convenience, when faced with the problem of irregular class boundaries and difficult to describe. At the same time, Most of the traditional classification algorithms can be boiled down to classification based on influence function.The experimental results of UCI data sets indicate that CBLIF, CBSIF, CBEIF and CBGIF can improve the classification performance and show significant advantage comparing to traditional classification methods. CBEIF can classify the data set with different degree of balance by adjusting the parameter in exponential influence function effectively. So, defining an appropriate function is helpful to understand a variety of classification methods and compare their relative advantages and disadvantages, and then to discover new classification method.