MagiNet: Mask-Aware Graph Imputation Network for Incomplete Traffic Data

TL;DR

This work tackles the challenge of imputing missing traffic data without resorting to pre-filling, which can inject noise and cause over-smoothing. It introduces MagiNet, a mask-aware graph imputation network built from an adaptive mask spatio-temporal encoder (AMSTenc) and a mask-aware spatio-temporal decoder (MASTdec) that leverage mask-aware attention and Chebyshev graph convolution to capture intrinsic spatio-temporal dependencies. Across five real-world datasets, MagiNet consistently outperforms both complete-data baselines with prefill and specialized incomplete-data imputation methods, achieving average improvements of 4.31% in RMSE and 3.72% in MAPE, with robustness across varying missing ratios. The approach advances practical ITS data completion by avoiding noise-prone pre-filling and better preserving spatial and temporal structure during imputation, with potential extensions to probabilistic imputation and large-scale deployment.

Abstract

Due to detector malfunctions and communication failures, missing data is ubiquitous during the collection of traffic data. Therefore, it is of vital importance to impute the missing values to facilitate data analysis and decision-making for Intelligent Transportation System (ITS). However, existing imputation methods generally perform zero pre-filling techniques to initialize missing values, introducing inevitable noises. Moreover, we observe prevalent over-smoothing interpolations, falling short in revealing the intrinsic spatio-temporal correlations of incomplete traffic data. To this end, we propose Mask-Aware Graph imputation Network: MagiNet. Our method designs an adaptive mask spatio-temporal encoder to learn the latent representations of incomplete data, eliminating the reliance on pre-filling missing values. Furthermore, we devise a spatio-temporal decoder that stacks multiple blocks to capture the inherent spatial and temporal dependencies within incomplete traffic data, alleviating over-smoothing imputation. Extensive experiments demonstrate that our method outperforms state-of-the-art imputation methods on five real-world traffic datasets, yielding an average improvement of 4.31% in RMSE and 3.72% in MAPE.

Figures (7)

• Figure 1: (a) Example of incomplete traffic data and corresponding ground truth from the Seattle dataset. (b) Performance comparison with and without pre-filling techniques to initialize missing values. Imputation performance with pre-filling techniques is significantly worse than without pre-filling. (c) Existing methods fail to capture inherent spatio-temporal correlations in incomplete traffic data, leading to an over-smoothing effect at dynamic missing positions, particularly noticeable between time steps 160 and 168.
• Figure 2: An overview of MagiNet, which consists of an adaptive mask spatio-temporal encoder (AMSTenc) and a mask-aware spatio-temporal decoder (MASTdec). The MASTdec stacks several spatio-temporal blocks (ST Block). Each ST Block combines a mask-aware spatio-temporal attention (MASTatt) module that calculates the mask temporal and spatial attention, an attention-based spatio-temporal aggregation module that aggregates spatial information using graph convolution and multi-scale temporal information using temporal convolution.
• Figure 3: Comparison of imputation curves between MagiNet and its variants for two snapshots: node #2 in Seattle and node #85 in METR-LA.
• Figure 4: Sensitivity analysis on METR-LA, Seattle, Chengdu, and Shenzhen datasets with respect to different missing ratio $r$ from 20% to 70%.
• Figure 5: Hyperparameter study on three key parameters of MagiNet: hidden size $\emph{h}$, spatio-temporal blocks $\emph{s}$, mask-aware spatial kernel size $\emph{k}$.

Theorems & Definitions (1)

• Definition 1