Graph Convolutional Network for Shared Bike Demand Forecasting: Time Series Modeling and Multi‑Task Learning

Shared two‑wheel vehicles are a crucial tool for short‑distance travel in cities. Because user rides cause uneven spatial distribution of vehicles, operators need to schedule bikes to match dynamic supply‑and‑demand conditions.

1. Introduction – The article explains the problem of estimating the amount of bikes to be dispatched at each location (riding‑in, riding‑out, and demand‑gap) and why accurate prediction is essential for improving user experience.

2. Basic Time‑Series Modeling – Feature engineering combines user, vehicle, road‑network, weather, time, and behavior sequence data. Linear models, tree models, and Facebook’s variants are mentioned as typical baselines. The goal is to predict future bike flow and the resulting demand gap at each point.

3. Multi‑Task Learning – Modeling each target (riding‑in, riding‑out, gap) separately ignores useful correlations. Multi‑task learning shares information among tasks, improving generalization and reducing over‑fitting.

4. Graph Convolutional Network (GCN) Exploration – The bike network is naturally represented as a graph (nodes = parking points, edges = travel relationships). Because the graph is non‑Euclidean, traditional convolution kernels cannot be applied directly. The article introduces spectral GCN based on the graph Laplacian and Fourier transform:

• The Laplacian matrix L = D – A (D: degree matrix, A: adjacency matrix) is symmetric positive‑semi‑definite, allowing eigen‑decomposition.

• Graph Fourier transform projects node features onto the eigenvectors of L, turning convolution into element‑wise multiplication in the spectral domain.

• Convolution in the graph domain is expressed as \(X * g_\theta = U g_\theta(\Lambda) U^T X\), where U are eigenvectors and \(\Lambda\) eigenvalues.

• Chebyshev polynomial approximation reduces computational cost, leading to the popular first‑order GCN propagation rule:

def call(self, layer_input):
    layer_input_f, DAD = layer_input
    output = tf.matmul(tf.matmul(DAD, layer_input_f), self.weight)
    if self.use_bias:
        output += self.bias
    act = self.activation(output)
    return act

The propagation aggregates normalized neighbor features (DAD) and applies a shared weight matrix, optionally adding bias and a non‑linear activation.

5. Advantages and Limitations of GCN – GCN captures complex topological information, shares parameters, and has locality (aggregates one‑hop neighbors). Stacking layers expands receptive fields, but too many layers may be unnecessary for short‑range bike travel. Limitations include handling directed graphs and uniform weighting of same‑order neighbors; extensions such as GAT and GraphSAGE address these issues.

6. Practical Application – The article provides a demo snippet for a basic GCN layer, discusses how to construct input feature matrix (N × d) and adjacency matrix (N × N) for the bike network, and mentions possible extensions (e.g., T‑GCN for temporal traffic prediction).

7. Conclusion – GCN effectively extracts spatial features from irregular bike‑sharing networks, but model choice should align with business stage and performance goals. The authors encourage further exploration and integration of graph‑based methods in demand‑prediction pipelines.