Research And Application Of Classification Method Based On Argumentation

Classification is one of the important techniques in data mining, machine learning and pattern recognition. It can map the examples into pre-defined categories according to the target function extracting from the training set. Target function is also called classification model. It is a speculation for the real classification model by using hypothesis during the course of machine learning. At present, one of the common classifications methods generally is to have a hypothesis, which gives consideration to training error and generalization error, as a classification model. This classification model can be used to evaluate the class properties of examples. The provided training set is usually credited with the property of consistency on the request of these methods, of which the classification effect will be easily disturbed by the noise or the absence of data. Besides, clusters have the possibility of being overlapped when class properties are defined in the light of clustering techniques. When the examples are covered by a few hypotheses, the traditional random selecting methods are not rigorous enough to resolve the conflicts.The classification method which is based on argumentation is a brand-new method differentiating from classical classification theories in recent years. It introduces the argumentation which is in the filed of artificial intelligence into classification. It can not only describe the classification method in a quantitative manner but also generate understandable classification reasons through the qualitative argumentation. Moreover, based on argumentation, the universal hypothesis rather the single hypothesis is the aim of this classification method. Therefore, it can effectively manage the inconsistency issues of training set data and have good-performing robustness.This paper puts forward a classification method on the basis of argumentation. First, it analyzes the classification techniques and argumentation theories and further explores the practicability of applying argumentation theory to classification. In addition, it provides a argumentation framework for classification which extends the argument in abstract argumentation framework into two types:classification argument and query argument and indicates the failure of hypothesis matching when the argumentation is being attacked. This framework also extends the relation in argumentation into four types: rebuttal, undercut, alliance and refusal; it quantizes the intensity of argumentation through confidence and takes it as an evidence for deciding whether the argumentation being attacked; it is used to formalize the classification argumentation process through an argumentation map. Finally, the argument evaluation algorithm raised in this paper will be used to analyze the state of argument which under the argumentation framework for classification. Four judgmental semantics: universal accepted, credulous accepted, refuse accepted and majority accepted, are presented on this basis, the aim of which is to select reasonable classification argument set as a decision-making gauge for classification.In order to validate the rationality of this method, this thesis sets up a argument support system which is exploited to classify two cases: medical diagnosis and court trail. The result shows that the method proposed in this thesis can explain the effectiveness of this classification result from the perspective of qualitative study, thus achieving the expected goal.