CGCNImp: a causal graph convolutional network for multivariate time series imputation.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series data frequently contains missing values due to sensor failures and data transmission issues.
• These missing values pose significant challenges for accurate data analysis and downstream applications.
• Existing imputation methods struggle with the complex correlations, temporal dependencies, and non-stationarity inherent in such data.

Purpose of the Study:

• To propose a novel model, CGCNImp, for effective multivariate time series imputation.
• To address the limitations of current methods in handling complex data characteristics.
• To improve the accuracy and reliability of time series data imputation.

Main Methods:

• CGCNImp integrates a correlation dependency module using neural Granger causality and a Graph Convolutional Network (GCN).
• A temporal dependency module employs an attention-driven Long Short-Term Memory (LSTM) network and a time lag matrix.
• Total variation reconstruction is utilized to address both missing values and noise.

Main Results:

• Empirical analyses were conducted on two real-world multivariate time series datasets.
• CGCNImp demonstrated superior imputation performance compared to existing state-of-the-art methods.
• The model effectively captures both cross-variable correlations and temporal dynamics.

Conclusions:

• CGCNImp offers a robust solution for imputing missing data in multivariate time series.
• The proposed model advances the field of time series imputation by effectively modeling complex dependencies.
• The findings suggest CGCNImp is a valuable tool for data preprocessing in various analytical tasks.