A depth K-means clustering method for time series data

Abstract

The invention discloses a depth K-means clustering method for time series data, which comprises the following steps of: obtaining a time sequence data set, preprocessing the time sequence data set, separating sample information and category information of data, and inputting the sample information of the data into a model; constructing a codec model, and adopting a convolution-deconvolution framework; Introducing the K-means clustering loss into a codec model to form a depth K-means clustering model facing the time series data, and fusing the extracted features with the clustering target; Training the constructed model by utilizing a back propagation algorithm, and guiding generation of a hidden layer state; Finally, performing k-means clustering based on the hidden layer state representation generated by the training process, and calculating the rand index index.. By fusing the extracted features with the clustering target, the generated hidden layer state not only can reconstruct theoriginal sample, but also is beneficial to the formation of a cluster-shaped structure, so that the clustering operation is better carried out, and the clustering precision is improved.

Description

Technical field [0001] The present invention relates to the technical field of time series clustering, in particular to a deep K-means clustering method for time series data. Background technique [0002] Clustering is a very common task in data mining. The purpose is to divide the data sample into a number of disjoint subsets, each of which forms a cluster. In particular, the distance between clusters should be as large as possible, and the distance within clusters should be as small as possible. Time series clustering is a type of clustering. [0003] Time series data are interesting because they are ubiquitous in various fields such as science, engineering, business, finance, economics, medical care, and government. The purpose of time series clustering is to identify groups with similar properties based on a given similarity measure. It is an important and useful technique for exploratory research on the characteristics of different groups in a given time series data set. ...

AI Technical Summary

Problems solved by technology

However, this type of time series clustering algorithm has several problems: first, it is more sensitive to noise and outliers; second, it is more sensitive to the shift in phase or amplitude of time series

However, the above type of methods are carried out in stages, that is, the features of the time series are first extracted, and then the clustering operation is performed

These two stages are independent and not combined, which may cause the extracted features not to be well suited for clustering

Images