Font Size: a A A

Research On Evolutionary Prediction Technology For Streaming Data

Posted on:2020-12-24Degree:MasterType:Thesis
Country:ChinaCandidate:Z W LiaoFull Text:PDF
GTID:2428330623951393Subject:Computer technology
Abstract/Summary:PDF Full Text Request
In recent years,with the development of information technology,the accumulation of data is rapidly developing,and the field of machine learning and data mining has also developed rapidly.But more and more data is appearing as streaming data.Different from traditional batch data,stream data has the characteristics of fast speed,large amount of data,non-reappearance and orderly,which makes traditional batch prediction technology less suitable for these stream data.Facing the various characteristics of data stream,the content of this paper is the evolutionary prediction technology for data stream.Incremental learning is mainly a technique for streaming data.When the source of data streams continuously,the original learning model is adjusted with the arrival of new data,and the model parameters also change.Concept drift is one of the characteristics of data streams,which is one of the biggest challenges in the field of data stream mining today.When the data stream gradually undergoes concept drift,the prediction performance of the learning model for the data stream gradually decreases due to changes in the data distribution.Therefore,we need to be able to detect whether it has a concept drift in time,then adjust and train the model,and then predict the dynamic data stream.This paper outlines the existing detection and solution methods for the occurrence of concept drift in data streams.Then,the research topics of this paper are proposed for the problems faced by the research institute.For dense data streams,this paper proposes a competitive integration algorithm based on infinite,high-speed,time-varying data streams that can incrementally learn data streams.The algorithm integrates two incremental model trees FIMT-DD,and the base model FIMT-DD is a time-varying learning data stream algorithm.In our algorithm,by learning two such incremental model trees,a better global or local base model is used as the final prediction model in the sliding window,which is more suitable for the new data stream to be reached..At the same time,the space complexity required by the algorithm is not very high.The detection and adaptation of local changes are achieved when concept drift occurs.Finally,the experimental results show that the algorithm can adapt well whether it is a smooth data stream or a non-stationary data stream.For the sparse data stream,this paper improves the current best performing onlinesparse data stream classification algorithm FTRL,which makes it well to deal with the problem of data stream concept drift.Because the sparse data stream drifts,the model parameters of the original algorithm FTRL are basically fixed,and the learning rate is very low.The common FTRL algorithm cann't learn a new concept.Therefore,in our improvement,the model spontaneously detects the concept drift of the data stream.If so,the FTRL model is adjusted and retrained to quickly adapt the FTRL to the new data stream.The simulated data stream shows that the improved model can adapt well to various concept drift data streams and has good robustness and stability.
Keywords/Search Tags:Data stream, data mining, concept drift, integration, sparse
PDF Full Text Request
Related items