Hybrid Generative-Discriminative Model Study Based On Attribute Division

Classification is an important task in the field of data mining, it has got people's attention because its wide application, the current classification model based on their modeling mechanism can be divided into two categories: generative model and discriminative classification model. Generative model generated the model within the main distribution of the classes, discriminative model the main distribution of between classes.Classification performance were compared in this paper on the generative and discriminative by experiments, and made a detailed comparison progressive mainly in the classification accuracy, the efficiency of modeling time, experimental conclusion: Classification of generative greatly influenced by the data distribution, but the modeling time is short, fast, for small data sets. Accuracy of discriminative classification is higher, but the modeling time is long and slow. The conclusion consistent with experimental results of Ng. Based on hybrid of generative and discriminative, generative and discriminative production hybrid model has become a research hotspot.The work based on hybrid model focused on supervised learning and semi-supervised learning areas, the study of hybrid classification model focused on two areas, one is the relationship between generative and discriminative, and the other is the build of hybrid model. Currently, There are three ways on construction of the hybrid model. They are partitioning based on attributes or data mixed classification model, the combination of generative and discriminative and learning from each other. Currently these mixed classification model has obtained better classification performance.Because the influence on the data distribution of generative, this paper presents a new hybrid classification model based on attribute division. Tests on data attributes, the attribute set into two subsets, respectively, classification models are established on the two subsets, and finally merge the two, applied it into mail classification,proved that the method fit the generative and discriminative classification method well, effectively improve the classification performance.