Research And Application Of Quantile Regression Based On Concept Drift Data Strea

Quantile regression,as a statistical prediction method that can reflect the overall condition distribution of the interpreted variable.It is also widely used in many fields due to its excellent characteristics such as being more robust and reliable in dealing with non normal distribution data with sharp peaks or thick tails,and it does not require distribution assumption for error terms.However,when making multilevel quantile regression predictions,the problem of quantile crossing is a common phenomenon,which will result in decreasing interpretability of the model.Besides,as the increasing of quantile regression levels,the model parameters for quantile regression will also increase,which will lead to increase the model forecasting difficulty.The rapid development of information science and technology has led to the emergence of data streams stacked in various fields.In these data streams,there is often a phenomenon of conceptual drift,in which the underlying distribution of data changes over time.Once conceptual drift occurs in the data stream,models based on historical data training cannot accurately depict future data.In order to obtain reliable predictions,it is necessary to understand the existing drift types in the data stream,and to know the drift mode of the current data instance during the modeling process.However,in practice,when a single piece of newly added data appears,it cannot be accurately classified as a certain type of drift.Therefore,this paper uses fuzzy logic to measure the membership of a single data instance belonging to different patterns,and embeds fuzzy membership into the learning process to handle the complexity and uncertainty of mixed drift.The artificial neural network has better fitting ability to nonlinear data.In response to the above situation,this paper builds a fuzzy clustering adaptive quantile regression model based on Back Propagation Neural Network,Recurrent Neural Network,Long Short Term Memory,Gated Recurrent Unit,and Temporal Convolutional Network,and combines the heavy-tailed quantile function to perform multilevel quantile collaborative prediction of concept drift data streams.In order to avoid the quantile crossing problem of multilevel quantile,this paper adopts a heavy-tailed quantile function with monotone property,which can not only depict the tail characteristics,but also reduce the number of parameters to be estimated in the model to a certain extent.In this model,the membership degree of data instances can be represented by the membership matrix obtained from fuzzy clustering,and then the membership matrix can be adapted to quantile regression parameters after synchronous training.This method can extract the valid information from the past to help predict the new data instances and realize the continuous tracking of the latest data patterns.Therefore,it can effectively avoid the risk of insufficient training due to the lack of new data,and take into account the online strategy or incremental strategy update function to improve prediction accuracy.In order to verify the superiority of the proposed fuzzy clustering adaptive quantile regression prediction model in processing concept drift data streams,multiple models are established as benchmark models for comparison.The validity of the proposed model is verified through three sets of real concept drift data streams.