PIMCST: Physics-Informed Multi-Phase Consensus and Spatio-Temporal Few-Shot Learning for Traffic Flow Forecasting

Accurate traffic flow prediction remains a fundamental challenge in intelligent transportation systems, particularly in cross-domain, data-scarce scenarios where limited historical data hinders model training and generalisation. The complex spatio-temporal dependencies and nonlinear dynamics of urban mobility networks further complicate few-shot learning across different cities. This paper proposes MCPST, a novel Multi-phase Consensus Spatio-Temporal framework for few-shot traffic forecasting that reconceptualises traffic prediction as a multi-phase consensus learning problem. Our framework introduces three core innovations: (1) a multi-phase engine that models traffic dynamics through diffusion, synchronisation, and spectral embeddings for comprehensive dynamic characterisation; (2) an adaptive consensus mechanism that dynamically fuses phase-specific predictions while enforcing consistency; and (3) a structured meta-learning strategy for rapid adaptation to new cities with minimal data. We establish extensive theoretical guarantees, including representation theorems with bounded approximation errors and generalisation bounds for few-shot adaptation. Through experiments on four real-world datasets, MCPST outperforms fourteen state-of-the-art methods in spatio-temporal graph learning methods, dynamic graph transfer learning methods, prompt-based spatio-temporal prediction methods and cross-domain few-shot settings, improving prediction accuracy while reducing required training data and providing interpretable insights. The implementation code is available at https://github.com/afofanah/MCPST.

💡 Research Summary

The paper tackles the persistent challenge of traffic flow forecasting under data‑scarce, cross‑domain conditions by introducing a novel framework called MCPST (Multi‑Phase Consensus and Spatio‑Temporal learning). The authors reconceptualize traffic prediction as a multi‑phase consensus problem, where three complementary physical dynamics—diffusion (modeling congestion propagation), synchronization (capturing rhythmic traffic patterns), and spectral structural analysis (encoding network topology)—are each modeled explicitly and then fused through an adaptive consensus mechanism.

In the diffusion module, traffic states are treated as a continuous diffusion process (∂u/∂t = κ∇²u) discretized on the graph Laplacian. Learnable diffusion coefficient κ and capacity C allow the model to adapt propagation speed and intensity to each city’s road characteristics. A source‑sink generator Q(F) conditions the diffusion on sensor features, providing a data‑driven initialization.

The synchronization module adopts a Kuramoto‑style formulation, learning node phases ϕₖ, intrinsic frequencies νₖ, and coupling strengths γₖ. This captures periodic phenomena such as rush‑hour peaks and enables a phase‑consistency loss that penalizes incoherent predictions across time.

The spectral module performs eigendecomposition of the graph Laplacian, extracting eigenvectors Ψ and eigenvalues Λ. The spectral gap g = λ₂ − λ₁ serves as a reliability indicator for poorly connected sub‑graphs, guiding the model to compensate for structural weaknesses.

Outputs from the three modules (F_diff, F_sync, F_spec) are combined by a reliability‑aware attention mechanism. The attention weights α_i are learned jointly with the prediction task, balancing each phase’s contribution based on its estimated confidence and the phase‑consistency regularizer. A multi‑scale temporal encoder—parallel LSTM and Transformer layers—processes historical sequences, while horizon‑specific heads produce forecasts for multiple future steps. Uncertainty is quantified via Bayesian dropout and neural consensus, yielding calibrated confidence intervals.

MCPST is trained using a meta‑learning scheme reminiscent of MAML. Each episode samples a support set (few labeled samples from a target city) and a query set. The support set adapts the base parameters θ to θ′ via a phase‑guided gradient step that incorporates the multi‑phase features as conditioning signals. The meta‑objective minimizes the expected sum of prediction loss and weighted phase‑consistency losses across episodes:

min_Θ E_S