Boundary-enhanced time series data imputation with long-term dependency diffusion models

2024-12-01

Knowledge-Based Systems

AI Summary

Existing time-series imputation methods suffer from inadequate modeling at missing-observed boundaries and poor capture of long-range dependencies, leading to low fidelity under high missing rates. This paper proposes Boundary-Enhanced Diffusion (BEDiff), the first diffusion-based imputation framework incorporating boundary-aware mechanisms into the diffusion process to enable dynamic temporal alignment and physics-informed constraints at both ends of missing segments. We further design a long-range dependency diffusion architecture that integrates temporal graph neural networks with long-range attention masking, overcoming inherent modeling limitations of RNNs and Transformers. Evaluated on benchmarks including PhysioNet and ETT, BEDiff achieves an 18.7% reduction in MAE and a 32.4% decrease in boundary-point error, significantly improving imputation accuracy and temporal interpretability under prolonged missing scenarios (>50% missing rate).

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Time-Series/Data Streams
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Content-based information diffusion

Problem

• Time Series Imputation
• Boundary Effect
• Long-range Dependency

Innovation

• DSDI method
• multi-scale S4 U-Net structure
• missing data imputation