扩散模型论文阅读 PriSTI: A Conditional Diffusion Framework for Spatiotemporal Imputation

0.Abstract

0.1 逐句翻译

Spatiotemporal data mining plays an important role in air quality monitoring, crowd flow modeling, and climate forecasting. However, the originally collected spatiotemporal data in real-world scenarios is usually incomplete due to sensor failures or transmission loss. Spatiotemporal imputation aims to fill the missing values according to the observed values and the underlying spatiotemporal dependence of them. The previous dominant models impute missing values autoregressively and suffer from the problem of error accumulation. As emerging powerful generative models, the diffusion probabilistic models can be adopted to impute missing values conditioned by observations and avoid inferring missing values from inaccurate historical imputation. However, the construction and utilization of conditional information are inevitable challenges when applying diffusion models to spatiotemporal imputation. To address above issues, we propose a conditional diffusion framework for spatiotemporal imputation with enhanced prior modeling, named PriSTI. Our proposed framework provides a conditional feature extraction module first to extract the coarse yet effective spatiotemporal dependencies from conditional information as the global context prior. Then, a noise estimation module transforms random noise to realistic values, with the spatiotemporal attention weights calculated by the conditional feature, as well as the consideration of geographic relationships. PriSTI outperforms existing imputa- tion methods in various missing patterns of different real-world spatiotemporal data, and effectively handles scenarios such as high missing rates and sensor failure. The implementation code is available at https://github.com/LMZZML/PriSTI.

时空数据挖掘在空气质量监测、人群流动建模和气候预测等方面发挥着重要作用。然而，现实世界中的原始时空数据通常由于传感器故障或传输损失而不完整。时空插值旨在根据观测值和它们背后的时空依赖关系填补缺失值。以前的主导模型自回归地填补缺失值，并遭受错误积累的问题。作为新兴的强大生成模型，扩散概率模型可以根据观测值条件地填补缺失值，并避免从不准确的历史插值中推断缺失值。然而，当应用扩散模型到时空插值时，构建和利用条件信息是不可避免的挑战。为了解决这些问题，我们提出了一个条件扩散框架，用于时空插值，并加强了先验建模，名为 PriSTI。我们提出的框架首先提供一个条件特征提取模块，以从条件信息中提取粗效但有效的时空依赖关系，作为全局背景先验。然后，一个噪声估计模块将随机噪声转换为现实值，利用条件特征计算时空注意力权重，并考虑地理关系。PriSTI 在各种现实世界的时空数据的缺失模式中优于现有的插值方法，并能够有效地处理高缺失率和传感器故障的情况。实现的代码可在 https://github.com/LMZZML/PriSTI 上获得。

Index Terms—Spatiotemporal Imputation, Diffusion Model, Spatiotemporal Dependency Learning

0.2

• 1.解决的时空数据挖掘问题，主要是补全缺失值的问题。
• 2.由于传统自回归的模型具有累积误差的问题，所以这里作者没有采用自回归的问题。

I. INTRODUCTION

1.1翻译

第一段（时空数据补全是常见的问题，有研究价值）

Spatiotemporal data is a type of data with intrinsic spatial and temporal patterns, which is widely applied in the real world for tasks such as air quality monitoring [1], [2], traffic status forecasting [3], [4], weather prediction [5] and so on. However, due to the sensor failures and transmission loss [2], the incompleteness in spatiotemporal data is a common problem, characterized by the randomness of missing value’s positions and the diversity of missing patterns, which results in incorrect analysis of spatiotemporal patterns and further interference on downstream tasks. In recent years, extensive research [1], [6], [7] has dived into spatiotemporal imputation, with the goal of exploiting spatiotemporal dependencies from available observed data to impute missing values.

时空数据是一种具有内在时空模式的数据，广泛应用于现实世界的任务中，如空气质量监测 [1],交通状况预测 [3],天气预测 [5] 等。然而，由于传感器故障和传输损失 [2],时空数据的不完整性是一个普遍的问题，其特征是缺失值位置的随机性和缺失模式的多样化，这会导致对时空模式的错误分析，并对后续任务产生进一步干扰。近年来，广泛的研究 [1], [6], [7] 已经深入探究了时空插值，旨在从可用的观测数据中挖掘时空依赖关系，以填补缺失值。

第二段（前人研究从历史值当中进行自回归，历史值本身就不要准确，这种方式会自然而然的扩大累积误差）

The early studies applied for spatiotemporal imputation usually impute along the temporal or spatial dimension with statistic and classic machine learning methods, including but not limited to autoregressive moving average (ARMA) [8], [9], expectation-maximization algorithm (EM) [10], [11], k- nearest neighbors (KNN) [12], [13], etc. But these methods impute missing values based on strong assumptions such as the temporal smoothness and the similarity between time series, and ignore the complexity of spatiotemporal correla- tions. With the development of deep learning, most effective spatiotemporal imputation methods [1], [7], [14] use the re- current neural network (RNN) as the core to impute missing values by recursively updating their hidden state, capturing the temporal correlation with existing observations. Some of them also simply consider the feature correlation [1] by the multilayer perceptron (MLP) or spatial similarity between different time series [7] by graph neural networks. However, these approaches inevitably suffer from error accumulation [6], i.e., inference missing values from inaccurate historical imputation, and only output the deterministic values without reflecting the uncertainty of imputation.

早期的时空插值研究通常使用统计和经典机器学习方法，例如自回归移动平均 (ARMA) [8], [9],期望最大化算法 (EM) [10], [11],k-最近邻 (KNN) [12], [13] 等，以沿着时间或空间维度进行插值。但是，这些方法基于强假设，如时间平滑性和时间序列之间的相似性，并忽略时空相关性的复杂性。随着深度学习的发展，最有效的时空插值方法 [1], [7], [14] 使用循环神经网络 (RNN) 作为核心，通过递归更新其隐藏状态来插值缺失值，并捕捉与现有观测相关的时间相关性。其中一些方法还简单地使用多层感知器 (MLP) 考虑特征相关性 [1] ,或使用图神经网络 (GNN) [7] 来考虑不同时间序列之间的空间相似性。然而，这些方法不可避免地遭受错误积累 [6],即从不准确的历史插值中推断缺失值，并仅输出确定性值，而不反映插值的不确定性。

第三段（扩散模型可以避免这种累积误差，前人利用扩散模型时候没有很好的利用更多周围信息和环境信息，而略微显得单薄了。）

More recently, diffusion probabilistic models (DPM) [15]– [17], as emerging powerful generative models with impressive performance on various tasks, have been adopted to impute multivariate time series. These methods impute missing values starting from randomly sampled Gaussian noise, and convert the noise to the estimation of missing values [18]. Since the diffusion models are flexible in terms of neural network architecture, they can circumvent the error accumulation prob- lem from RNN-based methods through utilizing architectures such as attention mechanisms when imputation, which also have a more stable training process than generative adversarial networks (GAN). However, when applying diffusion models to imputation problem, the modeling and introducing of the conditional information in diffusion models are the inevitable challenges. For spatiotemporal imputation, the challenges can be specific to the construction and utilization of conditional information with spatiotemporal dependencies. Tashiro et al. [18] only model temporal and feature dependencies by atten- tion mechanism when imputing, without considering spatial similarity such as geographic proximity and time series corre- lation. Moreover, they combine the conditional information (i.e., observed values) and perturbed values directly as the
input for models during training, which may lead to inconsistency inside the input spatiotemporal data, increasing the difficulty for the model to learn spatiotemporal dependencies.

较新的方法中，扩散概率模型 (DPM) [15]-[17] 作为在多种任务中表现出出色性能的新兴生成模型，被用于插值多元时间序列。这些方法从随机采样的高斯噪声开始，将其转化为缺失值的估计 [18]。由于扩散模型在神经网络架构方面非常灵活，因此它们可以通过使用注意力机制等架构，绕过基于 RNN 的方法的错误积累问题，同时比生成对抗网络 (GAN) 更具有稳定的训练过程。然而，当应用扩散模型到插值问题时，扩散模型中的条件信息和条件的建模和引入是不可避免的挑战。对于时空插值，这些挑战可能与时空依赖信息的构建和利用有关。Tashiro 等人 [18] 仅在插值时利用注意力机制建模时间和特征依赖关系，而不考虑地理邻近和时间序列相关等空间相似性。此外，它们在训练时直接将条件信息 (即观测值) 和扰动值组合作为模型的输入，这可能会使输入的时空数据内部出现不一致性，从而增加模型学习时空依赖关系的难度。

第四段（详细介绍实现细节）

To address the above issues, we propose a conditional diffusion framework for SpatioTemporal Imputation with en- hanced Prior modeling (PriSTI).We summarize the existing methods that can be applied to spatiotemporal imputation, and compare the differences between our proposed method and the recent existing methods, as shown in Figure 1. Since the main challenge of applying diffusion models on spatiotemporal imputation is how to model and utilize the spatiotemporal dependencies in conditional information for the generation of missing values, our proposed method reduce the difficulty of spatiotemporal dependencies learning by extracting condi- tional feature from observation as a global context prior. The imputation process of spatiotemporal data with our proposed method is shown in the right of Figure 1, which gradually transform the random noise to imputed missing values by the trained PriSTI. PriSTI takes observed spatiotemporal data and geographic information as input. During training, the observed values are randomly erased as imputation target through a specific mask strategy. The incomplete observed data is first interpolated to obtain the enhanced conditional information for diffusion model.

为了解决上述问题，我们提出了一种条件扩散框架，用于时空插值的增强先验建模 (PriSTI)。我们总结可用于时空插值的方法，并比较了我们提出的方法和最新方法之间的差异，如 Figure 1 所示。由于在时空插值中应用扩散模型的主要挑战是如何建模和利用条件信息中的时空依赖关系，以减少缺失值生成的难度，我们提出的方法通过从观察中提取条件特征作为全局上下文先验，从而减少时空依赖关系学习的难点。我们提出的时空数据插值过程在 Figure 1 的右侧所示，PriSTI 通过训练逐渐将随机噪声转化为插值缺失值。PriSTI 接受观测到的时空数据和地理信息作为输入。在训练期间，通过特定的掩膜策略随机擦除观测值，作为插值目标。观测数据首先进行插值以获得增强的条件信息，以便扩散模型使用。

For the construction of conditional in- formation, a conditional feature extraction module is provided to extract the feature with spatiotemporal dependencies from the interpolated informat