Research On Detection And Adaptive For Mixed Types Concept Drift

In real application scenarios,data often arrives continuously in the form of streams.Due to the dynamic changes of the environment,the stream data may have concept drift,that is,the concepts in the stream data will change over time.The phenomenon of concept drift is widely present in streaming data in various scenarios and fields,detecting and adapting to concept drift in time is the key to dealing with the problem of concept drift.There are many types of concept drift,and the existing concept drift detection and adaptive research mostly focus on a single type of concept drift,and it is difficult to adapt to a mixed type of concept drift.Real-streaming data usually contains mixed types of concept drift,which can accurately detect concept drift and determine its type,which can provide a good foundation for concept drift adaptation.An ideal classification model should be able to adapt to multiple types of concept drift at the same time.Therefore,this research mainly focuses on the detection and adaptation of the concept drift of mixed types.The main research work and innovation of this article are mainly reflected in the following two aspects:Aiming at the detection and type judgment of mixed-type concept drift,a method of detecting and type-judging mixed-type concept drift based on model accuracy and accuracy rate change rate is proposed.This method first trains the classifier model on the sample data set,and uses the model to predict new data;Then,set a threshold value pair to determine the state of concept drift;furthermore,compare the model prediction accuracy with the threshold pair to determine the state of concept drift;finally,determine the type of concept drift based on the length of time the accuracy rate changes.Experimental results show that the method can effectively detect concept drift on different data sets of single type and mixed type,and can accurately judge the type of drift at the same time.Aiming at the adaptive problem of the concept drift of mixed types,an adaptive framework that can deal with the concept drift of mixed types is proposed.This method first trains a classifier on the sample data to predict new data;secondly,based on the mixed type concept drift detection and type judgment method based on the model accuracy and the accuracy rate change rate,using sliding window method and integrated learning method to adapt to abrupt drift and gradual drift respectively.Then,when different types of concept drifts are detected,the two methods switch to each other to achieve the effect of dynamically adapting to the concept changes in the streaming data;Finally,the quality of the classifier is measured,and the corresponding update strategy is adopted for the classifiers of different quality.Experiment with six sets of artificial data with known specific drift conditions and four sets of real data with unknown drift conditions to compare with the existing concept drift adaptive algorithm.The experimental results show that the algorithm has better adaptive effect on mixed concept drift data.Aiming at the concept drift of two mixed types of sudden change and gradual change,this article adopts the method based on model accuracy for drift detection and type judgment,and combines sliding window method and integrated learning method to adapt to concept drift,and achieves good results.This research can provide new research ideas and new methods for the detection and adaptation of mixed-type concept drift.