A feed-forward dynamic ensemble classifier based on scenario characteristics
JU Chun-hu · Industrial Engineering and Engineering Management · 2013
Data mining techniques have been applied in many fundamental research domains such as retailing,stock market,telecommunications industry,and medicine. Data stream generated from the data in these industries are not stable and they change all the time. Moreover,these changes are unpredictable trigger dynamicity of target concepts which are generally known as concept drifts in the literature. Still,the relationships between hidden context and concepts are not clear. Modeling data flow which contains concept drift is one of core problems in the data mining field because the changeable target concept will reduce the accuracy of the model,and require that the corresponding decision model be revised to process the current inputted data. The models and algorithms used in the existing literature can be categorized into three groups:( 1) instance-based selection learning method,( 2) instance-based weighting learning method,and( 3) ensemble classification learning method( or learning with multiple concept descriptions). The base classifiers are used to reflect the current concept,and predict different classes of samples by integrating all the classification results. Ensemble classifier has been widely used to weaken the impact of concept drift on data stream classification models. When the predictive accuracy of one base classifier in these models is below the given threshold,the ensemble classifier begins to learn a new base classifier and replaces the old one to overcome the influence from the concept drift. However,the ensemble classifier starts to learn only when the accuracy of the base classifier is lower than the threshold. As a result,this may cause a certain lag from the identification of the current concepts for the ensemble classifier. This paper proposes a new method which adds scenario characteristics analysis to the ensemble classifier and adopts the information gain method to extract scenario characteristics. In addition,the threshold of the scenario characteristic is set dynamically to predict the occurrence of concept drift. When the variation of scenario characteristics exceeds the scenario threshold,the ensemble classifier is stimulated immediately to create a new base classifier rather than wait until the accuracy of the base classifier is below the given threshold,which makes the ensemble classifier capable of feed-forward learning. In this work,the proposed OCEC( Origin Characteristics Ensemble Classifier) model should be validated by several computational experiments because OCEC can reduce the integrated generalization error for mining concept drift data streams,and improve the effectiveness of concept drift detection.