CAST-CKT: Chaos-Aware Spatio-Temporal and Cross-City Knowledge Transfer for Traffic Flow Prediction

Traffic prediction in data-scarce, cross-city settings is challenging due to complex nonlinear dynamics and domain shifts. Existing methods often fail to capture traffic’s inherent chaotic nature for effective few-shot learning. We propose CAST-CKT, a novel Chaos-Aware Spatio-Temporal and Cross-City Knowledge Transfer framework. It employs an efficient chaotic analyser to quantify traffic predictability regimes, driving several key innovations: chaos-aware attention for regime-adaptive temporal modelling; adaptive topology learning for dynamic spatial dependencies; and chaotic consistency-based cross-city alignment for knowledge transfer. The framework also provides horizon-specific predictions with uncertainty quantification. Theoretical analysis shows improved generalisation bounds. Extensive experiments on four benchmarks in cross-city few-shot settings show CAST-CKT outperforms state-of-the-art methods by significant margins in MAE and RMSE, while offering interpretable regime analysis. Code is available at https://github.com/afofanah/CAST-CKT.

💡 Research Summary

The paper introduces CAST‑CKT, a novel framework for traffic flow forecasting that explicitly incorporates chaos theory into spatio‑temporal graph learning and cross‑city knowledge transfer, targeting few‑shot scenarios where only a handful of labeled observations are available in a new city. The authors first argue that traffic dynamics are fundamentally chaotic, exhibiting sensitivity to initial conditions, long‑range dependence, and multiple attractors. To capture these properties, they define a “chaos profile” for each city – a compact vector comprising Lyapunov exponents, Hurst exponents, sample entropy, correlation and box‑counting dimensions, recurrence statistics, and basic statistical descriptors (mean, variance, trend, seasonality). This profile serves both as a diagnostic of the current predictability regime (regular vs. chaotic) and as a quantitative alignment metric for cross‑city transfer.

The architecture consists of five main components: (1) chaos‑aware feature extraction that computes the chaos profile from minimal historical data; (2) parallel multi‑scale temporal encoding using down‑sampled LSTM streams (factors 1, 2, 4, 8) whose outputs are up‑sampled and fused; (3) a chaos‑conditioned attention mechanism where the query, key, and value projection matrices are generated by a lightweight conditioning network that takes the chaos profile as input, allowing the attention patterns to adapt to the current dynamical regime; (4) adaptive graph topology learning that combines node features and chaos features in a multi‑head attention to dynamically re‑estimate the adjacency matrix, thus handling heterogeneous sensor layouts and time‑varying spatial dependencies; and (5) multi‑horizon prediction with Bayesian dropout and ensemble techniques that produce calibrated uncertainty estimates whose width varies with the measured chaos level.

Training follows a meta‑learning paradigm. Each episode samples a support set and a query set from a distribution over cities and their chaos profiles. The model parameters are first adapted on the support set using the chaos profile (Adapt(θ, Kₛᵤₚₚ, C)), then evaluated on the query set. This meta‑objective encourages the network to learn a shared initialization that can be rapidly fine‑tuned for a new city with only a few samples, while the chaos profile guides the adaptation to the appropriate regime.

The authors provide a theoretical analysis showing that matching chaos profiles between source and target cities yields a bound on the transfer error, and they derive a generalisation bound that combines Rademacher complexity with a chaos‑consistency term. Empirically, CAST‑CKT is evaluated on four public traffic datasets (e.g., METR‑LA, PEMS‑BAY, Didi‑Chengdu, Taxi‑BJ) under strict few‑shot cross‑city settings (5–10 labeled samples in the target city). Compared with state‑of‑the‑art baselines such as ST‑GAT, GraphWaveNet, and MetaST, CAST‑CKT achieves 12–18 % lower MAE and RMSE on average, with up to 25 % error reduction in highly chaotic periods. Uncertainty calibration metrics (e.g., Expected Calibration Error) improve markedly, demonstrating that the model’s confidence intervals expand in chaotic regimes and contract when traffic is more predictable. Visualisations of attention weights and predicted uncertainty corroborate the interpretability claims: the model automatically detects regime shifts and adjusts its internal representations accordingly.

In summary, CAST‑CKT bridges chaos theory and modern graph‑based deep learning to enable robust, interpretable, and uncertainty‑aware traffic forecasting in data‑scarce, cross‑city contexts. The work opens avenues for real‑time chaos profile updating, integration of exogenous factors (weather, events), and deployment in large‑scale intelligent transportation systems.