Networked Time Series Imputation via Position-aware Graph Enhanced Variational Autoencoders

Area of Science:

• Data Science
• Machine Learning
• Network Science

Background:

• Multivariate time series (MTS) imputation is crucial for data analysis.
• Existing methods often neglect dynamic graph structures or assume complete graph information.
• Networked time series (NTS) present unique challenges due to changing graph topologies and missing edges.

Purpose of the Study:

• To address limitations in current MTS imputation methods for NTS.
• To propose a novel model capable of imputing missing values in both time series features and graph structures of NTS.
• To develop a method that leverages dynamic graph information for more accurate imputation.

Main Methods:

• Defined the problem of imputation over NTS with missing features and graph structures.
• Developed the PoGeVon model, utilizing a variational autoencoder (VAE) for imputation.
• Introduced a novel node position embedding based on random walk with restart (RWR) for enhanced expressiveness.
• Designed a multi-task learning decoder for reciprocal imputation of time series and graph structures.

Main Results:

• PoGeVon demonstrated superior performance compared to existing baseline methods.
• The RWR-based node embedding showed higher expressive power than traditional message-passing GNNs.
• The model effectively imputes missing values in both time series features and graph structures.

Conclusions:

• PoGeVon offers an effective solution for imputation in challenging NTS data.
• The proposed method successfully integrates dynamic graph information for improved imputation accuracy.
• This work advances the field of time series imputation by addressing complex networked data.