GraphWaveNetOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for GraphWaveNet (Graph WaveNet for Deep Spatial-Temporal Modeling).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → GraphWaveNetOptions<T>
For Beginners
GraphWaveNet achieves state-of-the-art traffic forecasting by combining:
The Key Insight: Traffic patterns have two types of dependencies: spatial (how one road affects nearby roads) and temporal (how patterns evolve over time). GraphWaveNet uses diffusion convolution for spatial modeling and dilated causal convolutions for temporal modeling.
What Problems Does GraphWaveNet Solve?
- Traffic speed/flow prediction on road networks
- Air quality forecasting across sensor networks
- Electricity load prediction across power grids
- Any spatio-temporal forecasting with graph structure
How GraphWaveNet Works:
- Adaptive Adjacency: Learns graph structure from node embeddings
- Diffusion Convolution: Bidirectional message passing on the graph
- Dilated Temporal Conv: WaveNet-style gated convolutions with exponentially growing dilation
- Skip Connections: Collects outputs from all layers for final prediction
GraphWaveNet Architecture:
- Node Embeddings: Learnable E1, E2 for adaptive graph A = softmax(ReLU(E1*E2^T))
- Diffusion Convolution: P^k * X * W for forward/backward random walks
- Gated TCN: (X * W_f + b_f) ⊙ σ(X * W_g + b_g) with dilated convolutions
- Skip Connections: Residual + skip from each layer to output
Key Benefits:
- No need for predefined graph structure (adaptive learning)
- Captures long-range temporal dependencies via dilated convolutions
- Bidirectional spatial propagation captures complex graph patterns
- Efficient parallel training unlike recurrent models
How It Works
GraphWaveNet combines graph convolution networks with WaveNet-style dilated causal convolutions to capture both spatial and temporal dependencies in time series data.
Reference: Wu et al., “Graph WaveNet for Deep Spatial-Temporal Graph Modeling”, IJCAI 2019.
https://arxiv.org/abs/1906.00121
Constructors
| Constructor | Summary |
|---|---|
GraphWaveNetOptions() |
Initializes a new instance of the GraphWaveNetOptions class with default values. |
GraphWaveNetOptions(GraphWaveNetOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseAdaptiveGraph { get; set; } |
Gets or sets whether to use adaptive graph learning. |
bool UsePredefinedGraph { get; set; } |
Gets or sets whether to also use a predefined adjacency matrix. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate for regularization. |
int DiffusionSteps { get; set; } |
Gets or sets the number of diffusion steps (K) for graph convolution. |
int DilationChannels { get; set; } |
Gets or sets the number of dilation channels. |
int EndChannels { get; set; } |
Gets or sets the number of end (output) channels. |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon. |
int LayersPerBlock { get; set; } |
Gets or sets the number of layers per temporal convolution block. |
int NodeEmbeddingDim { get; set; } |
Gets or sets the dimension of node embeddings for adaptive graph learning. |
int NumBlocks { get; set; } |
Gets or sets the number of temporal convolution blocks. |
int NumNodes { get; set; } |
Gets or sets the number of nodes in the graph. |
int NumSamples { get; set; } |
Gets or sets the number of samples for uncertainty estimation. |
int ResidualChannels { get; set; } |
Gets or sets the number of residual channels. |
int SequenceLength { get; set; } |
Gets or sets the sequence length (input time steps). |
int SkipChannels { get; set; } |
Gets or sets the number of skip connection channels. |