| With the development of computer and storage technology,data in various fields can be stored in the form of time series.How to process these data and dig out instructive information has gradually become a research hotspot in this field.In recent years,with the sustained development of machine learning,time series clustering algorithm has been proved to be effective in mining effective information in cloud computing and big data.However,due to the high dimension and correlation of time series data,many traditional clustering algorithms are difficult to directly mine the complex features of time series.The excellent performance of the convolution neural network in the task of time series classification proves that it can extract the features of time series well.However,due to the lack of labeled data,it is still a challenge to apply the convolutional neural network to time series clustering tasks.In this context,this paper mainly focuses on the time series clustering based on convolutional neural network,designs the algorithm and optimizes the model,and proposes the similarity measurement algorithm of single dimension time series based on convolutional neural network,the clustering algorithm of single-dimensional time series,and multidimensional time sequence clustering algorithm,which provides a theoretical basis for cloud computing and big data processing and analysis.At the same time,due to the development of information technology in the financial field,using neural networks to analyze financial data is a new direction for future development.Therefore,based on the proposed time series clustering algorithm,this paper analyzes the financial data and constructs the investment stock selection strategy to provide technical support for investors' investment decisions.The main contributions of this paper are as follows:(1)The traditional time series similarity measurement algorithm only considers the corresponding relationship between time points and cannot observe the trend change of the data from a global perspective.Based on the discovery that there is a positive correlation between the number of CNN output changes in the same direction and its similarity,a time series similarity measurement method is designed to analyze the differences between time series more effectively.The comparison results with two classic similarity measurement algorithms on 15 UCR data sets verify the effectiveness of the algorithm.(2)Based on the excellent performance of CNN in the field of time series classification,we draw on the idea of the DBSCAN algorithm,take the similarity between time series and the ranking as the standard,and design a priority aggregation with certain parameter settings.The algorithm for constructing labeled data sets with higher similarity data realizes the transformation of clustering problems into classification problems.So a CNN-based time series classification algorithm can be used to assist clustering and improve clustering performance.The effectiveness of the proposed algorithm is proved by extensive experiments on the UCR dataset with the prior art,and the experimental results show that the proposed algorithm achieves superior performance than other leading methods.The application of financial stock linkage analysis provides a reference for investment decision-making.(3)Aiming at the characteristics of correlation between different variables of multidimensional time series,the CNN-based clustering algorithm proposed above is improved.In the last layer of CNN,the features of each dimension are fused,and a new multi-dimensional time series clustering algorithm is proposed.The performance in 6 multi-dimensional time series data sets verifies the effectiveness of the algorithm.And on this basis,an investment stock selection strategy was designed,the optimal investment portfolio model was obtained,and a higher rate of return was obtained.Indirectly proves the effectiveness of the proposed multi-dimensional time series clustering algorithm in financial background. |
PDF Full Text Request