Abstract:In real-world scenarios, partial information losses of multivariate time series degrade the time series analysis. Hence, the time series imputation technique has been adopted to compensate for the missing values. Existing methods focus on investigating temporal correlations, cross-variable correlations, and bidirectional dynamics of time series, and most of these methods rely on recurrent neural networks (RNNs) to capture temporal dependency. However, the RNN-based models suffer from the common problems of slow speed and high complexity when dealing with long-term dependency. While some self-attention-based models without any recurrent structures can tackle long-term dependency with parallel computing, they do not fully learn and utilize correlations across the temporal and cross variable dimensions. To address the limitations of existing methods, we propose a novel so-called dual-branch cross-dimensional self-attention-based imputation (DCSAI) model for multivariate time series, which is capable of performing global and auxiliary cross-dimensional analyses when imputing the missing values. In particular, this model contains masked multi-head self-attention-based encoders aligned with auxiliary generators to obtain global and auxiliary correlations in two dimensions, and these correlations are then combined into one final representation through three weighted combinations. Extensive experiments are presented to show that our model performs better than other state-of-the-art benchmarkers on three real-world public datasets under various missing rates. Furthermore, ablation study results demonstrate the efficacy of each component of the model.(c) 2023 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

Time-delayed Multivariate Time Series Predictions.

TriD-MAE: A Generic Pre-trained Model for Multivariate Time Series with Missing Values

Generalizable Memory-driven Transformer for Multivariate Long Sequence Time-series Forecasting

DBT-DMAE: An Effective Multivariate Time Series Pre-Train Model under Missing Data

Biased Temporal Convolution Graph Network for Time Series Forecasting with Missing Values.

Wavelet-Temporal Neural Network for Multivariate Time Series Prediction

Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling

Temporal Tensor Transformation Network for Multivariate Time Series Prediction

Low-Rank Autoregressive Tensor Completion for Multivariate Time Series Forecasting

A Divide-and-Rule Combined Learning Method for Truly Multivariate Time Series Prediction

Temporal Dynamic Matrix Factorization for Missing Data Prediction in Large Scale Coevolving Time Series

Multivariate Time Series Modeling and Forecasting with Parallelized Convolution and Decomposed Sparse-Transformer

A hybrid framework for multivariate long-sequence time series forecasting

Multivariate Time Series Classification with Crucial Timestamps Guidance

High-Order Temporal Correlation Model Learning for Time-Series Prediction.

DSformer: A Double Sampling Transformer for Multivariate Time Series Long-term Prediction

Dual-branch cross-dimensional self-attention-based imputation model for multivariate time series

Sequence Attention for Multivariate Time Series Forecasting

A Multimodal Transformer for Live Streaming Highlight Prediction

Parallel Extraction of Long-term Trends and Short-term Fluctuation Framework for Multivariate Time Series Forecasting

TSARNet :Integrated Spatio-Temporal Attention Mechanism for Multivariate Time Series Prediction Deep Learning Framework