Dynamic graph-based bilateral recurrent imputation network for multivariate time series.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Multivariate time series imputation is crucial for data analysis.
• Existing graph neural networks (GNNs) often assume static correlations, which is unrealistic for dynamic real-world data.
• Dynamic correlations between variables change over time, necessitating advanced imputation techniques.

Purpose of the Study:

• To propose a novel dynamic graph-based bilateral recurrent imputation network (DGBRIN) for multivariate time series.
• To address the limitation of static correlation assumptions in existing GNN-based imputation methods.
• To accurately impute missing values in time series data with dynamically changing relationships.

Main Methods:

• Developed a dynamic adjacency matrix learning (DAML) module to capture localized, dynamic correlations within time series segments.
• Integrated temporal dependencies using an information fusion layer and Spearman rank correlation coefficient for dynamic adjacency matrices.
• Employed a hybrid graph-based bilateral recurrent network combining recurrent neural networks and graph convolutional networks for imputation.

Main Results:

• The proposed DGBRIN model demonstrated superior performance in multivariate time series imputation.
• Experiments on eight real-world datasets confirmed the model's effectiveness in handling dynamic correlations.
• The dynamic graph approach significantly improved imputation accuracy compared to static methods.

Conclusions:

• The DGBRIN model effectively addresses the limitations of static correlation assumptions in GNNs for time series imputation.
• Dynamic graph construction and hybrid recurrent-graph networks are promising for capturing complex temporal dependencies.
• The proposed method offers a robust solution for imputing missing values in dynamic multivariate time series data.