Research On Representation And Clustering Methods Based On Time-series Data

Time-series data is a kind of data recorded and indexed in time dimensional order,which is widely existed in various fields such as finance,Internet,bio-information,etc.In recent years,time-series data has become a hot research area.In recent years,temporal data mining has become a hot research topic,in which representation and clustering of temporal data are two important aspects.Representation is to represent temporal data into a form with specific structure or features to extract the information contained in it.Clustering is an unsupervised algorithm that aims to divide the unlabeled data set into different clusters.Cluster analysis can discover the underlying structure in the data and distinguish different patterns of the data.The represented features of the time-series data can be input as data sets into the regular clustering algorithm for the purpose of enhancing the clustering effect.The representation of time-series data and clustering are two closely related tasks,and the way of representing time-series data directly affects the effect of clustering.However,time-series data have complex features such as high latitude,high noise,periodicity and trend,which bring challenges to the processing of time-series data.The current research methods are affected by the high complexity of data noise and features,which restrict the representation capability of the time-series data and the performance of clustering.Deep learning has strong feature extraction and representation capability,however,the current clustering methods for time-series data are still being explored.In this thesis,a method based on fuzzy representation and deep clustering is studied for the characteristics of time-series data,specifically:(1)proposes a smoothing-based fuzzy representation(Smoothing fuzzy representation,FR)of time-series data,which uses smoothing to achieve noise reduction in order to cope with the data noise problem while ensuring that data features are preserved.The method fuzzy represents the temporal data,and the fuzzy logic can effectively deal with the uncertainty and incomplete information of the data and reduce the noise in the temporal data.Its computational simplicity does not lead to excessive growth in computation due to the increase in dimensionality and data magnitude.The fuzzy logic expression is derived from the abstraction of reality and is suitable for extracting features of time-series data with characteristics such as cycles and trends.The experiments show that the representation of time-series data using FR,input to the clustering algorithm,obtains a clustering effect superior to other time-series data representation methods,and the representation quality of the data achieves significant improvement.(2)A multi-feature fused asymmetric deep clustering model FFADC(Feature fusedasymmetric deep clustering,FFADC)for asymmetric time-series data is proposed,and the method can receive expressed data or raw data for clustering.The model uses recursive graphs to transform the time-series data,eliminating local noise and preserving global features.The multi-level,parallel computing deep neural network is computationally efficient in processing high-dimensional and high-order-of-magnitude time-series data.For the problem of complex feature extraction from time-series data,a recursive graph plus convolutional neural network,and a long and short-term memory network are used to extract different features for fusion,and an asymmetric variational self-encoder structure is used to increase the feature extraction capability.Clustering is embedded into the training process of the network,and the pattern recognition capability of the model is fully utilized for clustering.Experiments show that FFADC achieves significantly better clustering results than other algorithms on different time-series data.