Dynamic graph convolutional recurrent imputation network for spatiotemporal traffic missing data

Abstract

In real-world intelligent transportation systems, the spatiotemporal traffic data collected from sensors often exhibit missing or corrupted data, significantly hindering the development of traffic data research. Missing data imputation is a classic research topic that encompasses a wide range of methods. However, these methods are typically underdeveloped in two aspects: the dynamic spatial dependencies of the road network over time, and the information extraction and utilization of diverse data. In this study, we design a novel deep learning architecture – Dynamic Graph Convolutional Recurrent Imputation Network (DGCRIN) – as a tool to impute missing traffic data. The DGCRIN employs a graph generator and dynamic graph convolutional gated recurrent unit (DGCGRU) to perform fine-grained modeling of the dynamic spatiotemporal dependencies of road network. Additionally, an auxiliary GRU learns the missing pattern information of the data, and a fusion layer with a decay mechanism is introduced to fuse a diverse range of information. This architecture enables the DGCRIN to be highly adaptable to complex scenarios involving missing data. Extensive experiments on two datasets demonstrate the superiority of DGCRIN over multiple baseline models.

Introduction

Spatiotemporal traffic data collected from the real world via sensor devices are essential and fundamental for traffic research and applications [1]. By analyzing and mining traffic data, researchers can address a wide range of problems, including traffic speed prediction [2], traffic pattern recognition [3], and traffic data generation [4]. However, owing to the inherent unpredictability of the data collection and storage processes, the collected traffic information frequently contains missing data, which degrades model performance for downstream tasks if not handled properly. Therefore, it is necessary to perform imputation carefully on spatiotemporal traffic data.

Generally, the essence of traffic data imputation is to productively extract effective latent information, such as temporal correlations and spatial dependencies, from observed data to estimate the missing data. A wide range of methods have been developed to achieve this goal. Early approaches attempted to directly utilize statistical features, such as zeros, historical averages [5], and last observations [6], to fill in or simply eliminate the gaps in data. These rudimentary methods consider solely a site’s own historical data, whereas a more effective approach is to combine information from multiple sites. A typical KNN-based imputation method estimated missing data by averaging the known values of the $\mathit{k}$ neighbors. Recently, matrix- and tensor-based decomposition techniques [7], [8], [9] have exhibited great potential as tools to solve traffic data imputation problems. However, these methods rely on a global low rank while ignoring local spatiotemporal consistency. Therefore, these models may have certain limitations in capturing globally complex spatiotemporal dependencies. In contrast, deep learning-based approaches have exhibited extraordinary nonlinear modeling capabilities for various tasks [10], [11], [12], [13], [14], [15]. Recent studies have attempted to use neural networks (NN), such as recurrent neural networks (RNN) [16], convolutional neural networks (CNN) [17], and graph neural networks (GNN) [18], to handle data imputation tasks. Among them, the GNN-based models are highly effective in capturing spatial dependencies among irregular road networks than other methods. Although these methods have achieved some success in addressing the issue of missing data, two important research gaps remain.

Dynamic spatial dependencies: Most existing GNN-based methods model spatial correlations via predefined static graph structures based on geographic distance or road connectivity. Consequently, these methods maintain a constant graphical structure over time. However, traffic data often exhibit strong dynamic correlations in the spatiotemporal dimension, which may not be fully modeled by a static graph structure. Several recent studies have proposed the construction of dynamic graph structures to capture the dynamic dependencies of traffic data, and achieved promising performance in traffic prediction tasks [19], [20], [21]. However, this issue is rarely considered in the context of traffic data imputation. As a result, the construction of a dynamic graph that models dynamic dependencies, and provides more accurate and effective inference information for data reconstruction, remains a challenge.

Various complex temporal information: Traffic data are essentially time-series data collected through various physical devices at regular or irregular time intervals. In an environment prone to missing data, a variety of time-series datasets can be obtained from different perspectives. For example, according to the missing position and data collection timestamp, a masking matrix dataset and a time-lag matrix dataset can be generated. The former indicates the original data states that includes missing patterns, and the latter helps in studying the contributions of observations to the missing value estimates [22]. Both approaches are beneficial for the imputation task. Nevertheless, many prior studies [1], [23], [24] on the imputation of traffic data have neglected the analysis and utilization of such datasets. Therefore, the extraction and fusion of information from diverse and complex temporal data warrant further examination.

To address the aforementioned issues, a novel spatiotemporal deep learning approach for traffic data imputation called Dynamic Graph Convolutional Recurrent Imputation Networks (DGCRIN) is proposed in this study. A graph generator was developed to model dynamic spatial correlations and a dynamic graph convolutional gated recurrent unit (DGCGRU) was used to capture spatiotemporal dependencies. Furthermore, to extract more useful information from different types of data, we employed an auxiliary gated recurrent unit (GRU) to model the missing patterns of masking data, and then introduced a fusion layer with a decay mechanism to fuse information from disparate data. The missing values of different road segments at each timestamp are imputed in a bidirectional process using the spatial dependencies and temporal correlations learned from historical observations. To evaluate the effectiveness of the proposed model, we conducted experiments on two real-world traffic datasets, demonstrating the superior performance of our method over that of existing approaches. In addition, corresponding hyperparameter sensitivity and ablation experiments were also carried out.

Our main contributions can be summarized as follows:

• We propose a novel traffic data imputation framework to infer missing values in the spatiotemporal input data. The model can effectively model the dynamic spatiotemporal characteristics of incomplete traffic data by considering the road network’s dynamics and diverse temporal information including data missing patterns and observation slot, to achieve more accurate estimation.

• We propose a dynamic graph generation technique to perform fine-grained modeling of the spatial correlations between road network nodes with incomplete traffic data via using the recurrent generated imputation and historical information.

• We demonstrate the capabilities and advantages of the proposed model by applying it to two real-world traffic datasets under three types of missing patterns. Compared with the baselines, our model significantly reduced the imputation error while achieving a higher imputation accuracy.

The remainder of this paper is organized as follows. In Section 2, we first review existing studies pertaining to traffic data imputation by category. The imputation problem is formulated in Section 3, and the methodology is described in Section 4. The experiments conducted within this study are presented in Section 5. Finally, we conclude the paper in Section 6.