| Nowadays,the increasing complexity of data leads to the need for high-quality classification systems.In view of the advantages of ensemble learning in improving classification performance,the classification system based on ensemble learning and designed by independent innovation method has important research and application value.In the current ensemble learning methods,the problem of rigorous selection and system simplification is often neglected when integrating different models.For the C4.5 model,which is suitable for homogeneous integration,there is no comprehensive consideration of the threshold sequence of attributes,the impact of attributes on the target and the deviation of gain rate,thus affecting the accuracy.In view of the above problems,this thesis studies and designs an integrated learning classification system.Specific research work is as follows:(1)Aiming at alleviating the problems in heterogeneous ensemble method,a method based on ranking selection measure and adaptive weight setting is proposed.According to different classification models,the cross-validation method combined with composite index measurement method constructed based on the error entropy and complementarity of design,which can pick out two models with the highest ensemble priority,is adopted in this method.The composite index measurement method based on combination index can measure the ensemble priority and obtain the sort index of the remaining models,according to that,the final weight values can be found for different models for ensemble processing through adaptive weight setting means.The experimental results show that,the classification evaluation indexes of the proposed method are improved compared with other classification models which verifies the effectiveness of this approach.(2)Aiming at alleviating the problems in homogeneous ensemble method,a combinatorial ensemble method based on improved C4.5 algorithm is proposed.On the one hand,when dealing with the continuous attributes in C4.5 algorithm,the sequence range of candidate segmentation thresholds is narrowed by sampling and threshold supplementation near the boundary of attribute intervals corresponding to different classes and the attribute effects on target classification can be taken into consideration based on simultaneously using the information conducive to attribute selection and local influence information of attributes in information gain by measuring the relation between attributes and objects through Standardized Euclidean Distance and adding the influence degree of different components on different subsets under the same attribute to the subset entropy value calculation.On the other hand,when dealing with split information,the effects of redundancy and gain rate deviation between attributes can be mitigated by calculating the average Gini index of other attributes to test attributes and adding correction factor at the same time.The final effectiveness of this method is validated by the experimental results which shows that the proposed method can effectively improve the classification accuracy.(3)A classification system dominated by the integrated methods mentioned above which also includes the common classification algorithms is designed and developed,which is mainly composed of three functional modules,pretreatment,classification and evaluation.The classification module,as the core part,achieves three functions: heterogeneous integration,homogeneous integration and comprehensive classification.It is proved that the system has practical application value for that the advantages in dealing with data classification tasks has been shown in the final testing results. |
