Coupled Layer-wise Graph Convolution for Transportation Demand Prediction

• Precise regional demand forecasts can support advance resource allocation, rebalancing, and improved daily travel service.Source: Introduction · p. 1
• Existing transportation prediction research progressed from regional statistical analysis to deep learning models for spatio-temporal correlations.Source: Introduction · pp. 1–2
• Most GCN approaches use heuristically fixed adjacency matrices and do not adequately model hierarchical spatial dependence.Source: Introduction · p. 2
• CCRNN introduces CGC with self-learned adjacency matrices that vary across layers to extract multi-level spatial dependence adaptively.Source: Introduction · p. 2
• A layer-wise coupling mechanism links upper- and lower-level topological structures while reducing computational costs.Source: Introduction · p. 2
• A unitary sequence-to-sequence framework integrates multi-level spatial hidden states with a gated recurrent unit for prediction.Source: Introduction · p. 2

• Early transportation prediction methods combined data mining with empirical statistics but handled limited regions and missed simultaneous spatio-temporal correlations.Source: Transportation Demand Prediction · p. 2
• GCNs generalize convolutional neural networks to non-Euclidean data and have attracted increasing research attention.Source: Graph Convolutional Network · p. 2
• Graph convolutional networks are commonly categorized into spatial-based and spectral-based methods.Source: Graph Convolutional Network · p. 2
• Spatial methods aggregate neighboring features, whereas spectral methods focus on graph-signal filtering and filter structure.Source: Graph Convolutional Network · p. 3
• Existing methods generally leverage an initial adjacency matrix or its variants, making efficient and accurate multi-level dependence capture difficult.Source: Graph Convolutional Network · p. 3

• Station-based transportation maps stations to graph nodes, while station-less demand locations are treated as gathering around discrete places.Source: Station-level Demand Prediction · p. 3
• Given historical graph signals, the adjacency matrix A is learned with a data-driven mapping to complete the transportation graph.Source: Station-level Demand Prediction · p. 3
• Transportation demand prediction maps P historical graph-signal steps to the next Q forecast steps.Source: Station-level Demand Prediction · p. 3
• The normalized adjacency matrix is defined as  = D^-1A, where D contains node degrees.Source: Graph Convolutional Network · p. 3
• Graph propagation uses a K-step diffusion process with filter parameters θ applied to input signals X.Source: Graph Convolutional Network · p. 3
• Conventional graph convolutions use an initial adjacency matrix and its powers, whereas CGC learns varying adjacency matrices across layers for hierarchical representations.Source: Graph Convolutional Network · p. 3

Adjacency Matrix Generation

• CCRNN generates adjacency matrices from compact station representations obtained by decomposing graph-signal data with SVD.The station-wise representation Xs captures compact, high-level station features, and pairwise similarity determines adjacency edge weights.Source: Adjacency Matrix Generation · pp. 3–4

Transportation Graph Construction

• The framework uses a unified graph formalization for station-based and station-less transportation.Station nodes represent fixed infrastructure, while concentrated pickup and drop-off locations can form virtual stations.Source: Coupled Layer-wise Graph Convolution · p. 4

Coupled Layer-wise Graph Convolution

• CGC assigns different adjacency matrices to different graph-convolution layers to capture multi-level spatial dependence.The layer outputs are recursively connected, with A(m) varying across layers.Source: Coupled Layer-wise Graph Convolution · p. 4
• A coupled mapping constructs upper-level adjacency matrices from lower-level matrices and reduces parameterization through shared low-dimensional node embeddings.The mapping parameters are shared between E1 and E2, with each coupled mapping using L × (L + 1) parameters.Source: Coupled Layer-wise Graph Convolution · pp. 4–5
• The first-layer adjacency matrix is initialized from the original graph structure and optimized with stochastic gradient descent to discover spatial relationships.This replaces purely random initialization while retaining end-to-end adaptation.Source: Coupled Layer-wise Graph Convolution · pp. 4–5

Temporal Dependence Modeling

• Attention aggregates representations from all graph-convolution layers before feeding the result into CCGRU for sequence-to-sequence forecasting.The recurrent unit uses reset and update gates, while scheduled sampling supports multi-step prediction.Source: Multi-level Aggregation · p. 5, Temporal Dependence Modeling · p. 5

Experimental Setup

• Experiments use NYC Citi Bike and NYC Taxi datasets, comparing CCRNN with seven forecasting baselines.The baselines include HA, XGBoost, FC-LSTM, DCRNN, STGCN, STG2Seq, and Graph WaveNet.Source: Datasets · p. 5, Baselines · p. 6
• The study retains 250 bike stations and clusters taxi orders into 266 virtual stations within an 8.42km × 14.45km region.Demand is modeled with half-hour time steps, using the first two of the last four weeks for validation and the last two for testing.Source: Experimental Setup · p. 6

Main Results

• CCRNN achieves the best performance on all evaluation metrics across both NYC datasets.Models are evaluated on multi-step outputs, and each baseline is trained 10 times for an average result.Source: Main Results · p. 6
• 13.85% and 26.85% lower RMSE than Graph WaveNet on the two datasets indicates more accurate and efficient transportation-demand dependency capture.Graph WaveNet remains competitive because it learns an adaptive adjacency matrix.Source: Main Results · p. 7

Multi-step Demand Prediction

• Across selected 0.5-, 2.5-, 4.5-, and 6.0-hour horizons, CCRNN achieves the lowest RMSE and highest PCC despite sequence-to-sequence performance decline.Graph WaveNet is stable across horizons but has lower short-term accuracy, while DCRNN declines as forecast time increases.Source: Main Results · p. 7, Figure 3 · p. 7

Ablation Study

• The complete CCRNN outperforms its ablated variants, while distance- and PCC-initialized variants still outperform all Table 1 baselines.Randomly initializing the embedding matrices produces PCC below 0.1, indicating that proper adjacency initialization is necessary for training.Source: Ablation Study · p. 7, Table 2 · p. 7

Conclusion

• CCRNN uses CGC with self-learned adjacency matrices that vary across layers to capture multi-level spatial dependence.A layer-wise coupling mechanism bridges upper- and lower-level graph structures and reduces model parameter scale.Source: Conclusion · p. 7
• A multi-level aggregation module weights extracted representations before a unitary network fuses them for final predictions.The paper frames this architecture as integrating spatial representations with recurrent prediction.Source: Conclusion · p. 7
• Experiments on real-world taxi and sharing-bike datasets report state-of-the-art results for CCRNN.The paper presents the model as a new perspective on graph convolution with layer-wise adjacency matrices.Source: Conclusion · p. 7