TSI-GNN: Extending Graph Neural Networks to Handle Missing Data in Temporal Settings

Area of Science:

• Data Science
• Machine Learning
• Graph Representation Learning

Background:

• Missing data is prevalent in time-series observations, often requiring imputation methods with strong distributional assumptions.
• Existing imputation techniques may not adequately consider temporal dynamics or generalize well to new time-dependent data without retraining.

Purpose of the Study:

• To develop a novel approach for data imputation that effectively incorporates temporal information into bipartite graphs.
• To address limitations of current methods that often overlook temporality or require retraining for time-series data.

Main Methods:

• Proposed a novel method, temporal setting imputation using graph neural networks (TSI-GNN), leveraging joint bipartite graphs and graph representation learning.
• Utilized observation nodes and edges with temporal information in message passing for learning embeddings and informing imputation.
• Captured sequence information within a graph neural network's aggregation function.

Main Results:

• TSI-GNN demonstrated improved data representation at 30% and 60% missing rates on benchmark datasets.
• Performance was particularly enhanced when using a nonlinear model for downstream prediction tasks on regularly sampled datasets.
• The method proved competitive with existing temporal imputation techniques across various scenarios.

Conclusions:

• Incorporating temporal information into bipartite graphs via TSI-GNN enhances data representation for missing data imputation.
• TSI-GNN offers a robust and competitive approach for handling missing data in time-series contexts.
• The study provides insights into managing the model size of TSI-GNN for practical applications.