Overview
This research introduces a novel Multi-View Coupled Tensor Decomposition (MVCTD) model designed for online adaptive traffic prediction. The model addresses the challenges posed by imperfect sensing conditions in intelligent transportation systems, particularly missing observations and anomalous disturbances, which complicate continuous forecasting. MVCTD leverages coupled tensor decomposition to construct a structured latent forecasting space, enabling the joint modeling of shared spatial structures across different traffic views (e.g., speed, flow, occupancy) and view-specific temporal dynamics. A key feature of MVCTD is the incorporation of group sparse regularization, which aims to capture correlated abnormal responses stemming from real traffic anomalies, thereby mitigating their influence on forecast accuracy. For streaming deployment, the model incorporates an iterative refinement process focused solely on the current latent tensor, while other model variables are updated efficiently through lightweight closed-form steps based on summarized historical information, avoiding re-optimization over full historical data.
Research Context
Accurate online traffic prediction is identified as an essential component for intelligent transportation systems. The continuous nature of forecasting in these systems presents inherent difficulties due to imperfect sensing conditions. Specific challenges highlighted include the presence of missing observations and anomalous disturbances within traffic data. The problem is further compounded when prediction systems rely on a single traffic view, which may limit their robustness to these imperfections. The development of methods that can address these issues, especially in an online adaptive manner, is crucial for improving the reliability and utility of traffic forecasting systems.
Approach
The core of the proposed method is the Multi-View Coupled Tensor Decomposition (MVCTD) model. This model is engineered to operate on imperfect multi-view observations, which can include diverse data streams such as speed, flow, and occupancy. The methodological framework integrates several key components:
- Coupled Tensor Decomposition: This technique is utilized to establish a structured latent forecasting space. Within this space, the model simultaneously captures shared spatial structures that exist across different traffic views and distinct temporal dynamics specific to each view.
- Group Sparse Regularization: To enhance robustness against data imperfections, a group sparse regularization technique is introduced. The purpose of this regularization is to identify and capture correlated abnormal responses that arise from actual traffic anomalies. By addressing these anomalous responses, the model aims to reduce their detrimental impact on the accuracy of forecasts.
- Online Adaptive Mechanism: For practical streaming deployment, MVCTD employs an efficient update strategy. Iterative refinement is applied exclusively to the current latent tensor. Concurrently, the remaining model variables are updated through lightweight closed-form steps. This update mechanism relies on summarized historical information, which is a design choice intended to circumvent the need for repeated, computationally intensive optimization over the entire historical data sequence.
Findings
Experiments conducted on real-world traffic datasets yielded several key findings regarding the MVCTD model:
- MVCTD achieved accurate forecasts.
- The model demonstrated a favorable runtime performance.
- These results were observed even under conditions characterized by severe missingness in the data.
- The experimental outcomes confirm the suitability of MVCTD for online traffic prediction applications.
Why This Matters
The demonstrated ability of MVCTD to provide accurate forecasts with favorable runtime under severe missingness is critical for intelligent transportation systems. Its suitability for online traffic prediction, even with imperfect sensing conditions, supports the continuous operation of these systems.