SUSTeR: Sparse Unstructured Spatio Temporal Reconstruction on Traffic Prediction

TL;DR

SUSTeR tackles traffic forecasting under extreme sparsity and spatial irregularity by transforming sparse observations into a dense hidden graph through a residual, context-aware encoding. A self-adaptive adjacency learned from node embeddings enables spatio-temporal correlation mining via a STGCN-based backbone, and a decoder predicts values at arbitrary query locations. The approach yields strong performance in very sparse regimes (up to $99.9\%$ missing data), with faster training and fewer parameters than baselines, and demonstrates robustness to reduced training data. This framework broadens applicability to moving sensors and regions lacking fixed road-network priors, with potential extensions to domains like oceanography where observations are sparse and mobile.

Abstract

Mining spatio-temporal correlation patterns for traffic prediction is a well-studied field. However, most approaches are based on the assumption of the availability of and accessibility to a sufficiently dense data source, which is rather the rare case in reality. Traffic sensors in road networks are generally highly sparse in their distribution: fleet-based traffic sensing is sparse in space but also sparse in time. There are also other traffic application, besides road traffic, like moving objects in the marine space, where observations are sparsely and arbitrarily distributed in space. In this paper, we tackle the problem of traffic prediction on sparse and spatially irregular and non-deterministic traffic observations. We draw a border between imputations and this work as we consider high sparsity rates and no fixed sensor locations. We advance correlation mining methods with a Sparse Unstructured Spatio Temporal Reconstruction (SUSTeR) framework that reconstructs traffic states from sparse non-stationary observations. For the prediction the framework creates a hidden context traffic state which is enriched in a residual fashion with each observation. Such an assimilated hidden traffic state can be used by existing traffic prediction methods to predict future traffic states. We query these states with query locations from the spatial domain.

Figures (6)

• Figure 1: Showing the novel problem statement applied to traffic prediction use case. Multiple unstructured observations from the past are used to reconstruct a hidden traffic state from which a full traffic state is forecast with a set of query locations.
• Figure 2: Framework of SUSTeR with an observation encoding, a residual architecture for hidden traffic state reconstruction with a variable amount of observations and a decoding from the dense hidden traffic state into the original space.
• Figure 3: Mean absolute error for SUSTeR and the baselines D2STGNN and STGCN with both modifications as average over 5 runs. The dropout rates are plotted in logarithmic scaling for better visualization.
• Figure 4: The mean absolute error (MAE) for the PEMS-BAY dataset evaluated on different dropout rates.
• Figure 5: Training only on a fraction of the training set while keeping validation and test set the same. Executed with a 99% dropout rate.