Core Concepts
Revisiting time series imputation from a causal perspective to discover and block confounders, leading to the development of Casper, a novel Causality-Aware Spatiotemporal Graph Neural Network.
Abstract
1. Introduction
Importance of spatiotemporal time series in understanding human activities.
Challenges posed by missing data due to various failures.
Shift towards deep learning methods for time series imputation.
2. Methodology
Introducing the concept of Structure Causal Model (SCM) for understanding causal relationships.
Frontdoor adjustment to eliminate backdoor paths caused by unknown confounders.
Description of Casper architecture with Spatiotemporal Causal Attention (SCA) and Prompt Based Decoder (PBD).
3. Framework Analysis
Theoretical proof of convergence for the causality indicator ๐ in SCA.
Complexity analysis highlighting the efficiency of Casper's components.
4. Experiments
Evaluation on real-world datasets AQI, METR-LA, PEMS-BAY, and AQI-36.
Comparison with traditional statistical methods, early deep learning models, and recent deep learning models.
Stats
่่
ใใฏใๆฐใใCausality-Aware Spatiotemporal Graph Neural Network๏ผCasper๏ผใๅฐๅ
ฅใใพใใใ