This paper focuses on the multivariate time series imputation problem using
deep neural architectures. The ubiquitous issue of missing data in both
scientific and engineering tasks necessitates the development of an effective
and general imputation model. Leveraging the wisdom and expertise garnered from
low-rank imputation methods, we power the canonical Transformers with three key
knowledge-driven enhancements, including projected temporal attention, global
adaptive graph convolution, and Fourier imputation loss. These task-agnostic
inductive biases exploit the inherent structures of incomplete time series, and
thus make our model versatile for a variety of imputation problems. We
demonstrate its superiority in terms of accuracy, efficiency, and flexibility
on heterogeneous datasets, including traffic speed, traffic volume, solar
energy, smart metering, and air quality. Comprehensive case studies are
performed to further strengthen the interpretability. Promising empirical
results provide strong conviction that incorporating time series primitives,
such as low-rank properties, can substantially facilitate the development of a
generalizable model to approach a wide range of spatiotemporal imputation
problems.

这篇论文研究了使用深度神经网络架构解决多元时间序列插补问题。通过利用低秩插补方法的经验和专长，我们为传统的 Transformer 模型添加了三项知识驱动的增强，包括投影时间注意力、全局自适应图卷积和 Fourier 插补损失。这些面向任务的归纳偏置利用了不完整时间序列的内在结构，使得我们的模型适用于多种插补问题。我们通过在包括交通速度、交通流量、太阳能、智能电表和空气质量等异构数据集上进行全面的案例研究来进一步加强其可解释性，并通过有希望的实证结果有力地证明了将低秩属性等时间序列原始特征纳入模型可以大幅促进通用化模型的开发，从而解决广泛的时空插补问题。