Abstract:Unsupervised learning of time series data, also known as temporal clustering, is a challenging problem in machine learning. Here we propose a novel algorithm, Deep Temporal Clustering (DTC), to naturally integrate dimensionality reduction and temporal clustering into a single end-to-end learning framework, fully unsupervised. The algorithm utilizes an autoencoder for temporal dimensionality reduction and a novel temporal clustering layer for cluster assignment. Then it jointly optimizes the clustering objective and the dimensionality reduction objec tive. Based on requirement and application, the temporal clustering layer can be customized with any temporal similarity metric. Several similarity metrics and state-of-the-art algorithms are considered and compared. To gain insight into temporal features that the network has learned for its clustering, we apply a visualization method that generates a region of interest heatmap for the time series. The viability of the algorithm is demonstrated using time series data from diverse domains, ranging from earthquakes to spacecraft sensor data. In each case, we show that the proposed algorithm outperforms traditional methods. The superior performance is attributed to the fully integrated temporal dimensionality reduction and clustering criterion.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is unsupervised learning in time - series data, especially temporal clustering. Time - series data has a complex structure and shows significant variation characteristics in different application fields, such as time scale, dimension, etc. In addition, time - series data in practical applications often has time - break points and high - frequency noise, which makes it very challenging to use traditional clustering techniques to process time - series data. The paper proposes a new algorithm - Deep Temporal Clustering (DTC), aiming to naturally integrate dimension reduction and time clustering through an end - to - end learning framework to achieve completely unsupervised learning of time - series data. The key of the DTC algorithm is to use a trainable network (a deep auto - encoder network is selected in this paper) to transform time - series data into a low - dimensional latent space and combine it with a new time - clustering layer. This method can not only effectively capture the complex features in the time - series, but also improve the clustering accuracy by jointly optimizing the dimension - reduction objective and the clustering objective. Specifically, the DTC algorithm achieves this goal through the following steps: 1. **Dimension Reduction**: First, use a one - dimensional convolutional neural network (1D CNN) to extract short - term features, and then further reduce the data dimension through a bidirectional long - short - term memory network (Bi - LSTM) and learn the temporal connections between waveforms on different time scales. 2. **Clustering**: Based on the latent representation of Bi - LSTM, find one or more spatio - temporal dimensions through a non - parametric clustering method, so that the data can be divided into two or more categories. 3. **Visualization**: Provide a method to generate a heat map to visualize the most informative areas in the time - series for clustering decisions, which is a function that traditional clustering algorithms do not have. The paper proves through experiments on multiple time - series datasets from different fields that the DTC algorithm has better performance compared to traditional methods (such as k - Shape and complete - link - based hierarchical clustering). This is mainly due to the fact that the DTC algorithm can simultaneously optimize the reconstruction loss and the clustering loss, thus maintaining the time continuity and complex structure of the data while reducing the dimension.

Deep Temporal Clustering : Fully Unsupervised Learning of Time-Domain Features

Deep Spatiotemporal Clustering: A Temporal Clustering Approach for Multi-dimensional Climate Data

Deep Clustering and Visualization for End-to-End High-Dimensional Data Analysis.

Deep Temporal Contrastive Clustering

Recurrent Deep Divergence-based Clustering for simultaneous feature learning and clustering of variable length time series

Deep Temporal Graph Clustering

Deep Time Series Sketching and Its Application on Industrial Time Series Clustering

End-to-end deep representation learning for time series clustering: a comparative study

An autoencoder-based deep learning approach for clustering time series data

Unsupervised Representation Learning for Time Series with Temporal Neighborhood Coding

A Deep Trajectory Clustering Method Based on Sequence‐to‐sequence Autoencoder Model

Deep Time-Series Clustering: A Review

Self-Supervised Time Series Clustering With Model-Based Dynamics

Hybrid Ensemble Deep Graph Temporal Clustering for Spatiotemporal Data

Bridging the Gap: A Decade Review of Time-Series Clustering Methods

Hierarchical Clustering using Auto-encoded Compact Representation for Time-series Analysis

Deep Temporal Joint Clustering for Satellite Image Time-Series Analysis

Unsupervised Spatio-temporal Latent Feature Clustering for Multiple-object Tracking and Segmentation

Unsupervised Visual Time-Series Representation Learning and Clustering

FeatTS: Feature-based Time Series Clustering

Deep Feature Unsupervised Domain Adaptation for Time-Series Classification