MTGNNOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for MTGNN (Multivariate Time-series Graph Neural Network).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → MTGNNOptions<T>
For Beginners
MTGNN is unique because it LEARNS how variables are connected:
The Key Insight: Unlike other graph models that require you to define the graph structure upfront, MTGNN automatically discovers which time series influence each other through an adaptive graph learning module. This is powerful when relationships are unknown.
What Problems Does MTGNN Solve?
- Traffic prediction when road network is complex or unknown
- Multivariate financial forecasting with unknown correlations
- Sensor networks where dependencies change over time
- Any multivariate time series where inter-variable relationships matter
How MTGNN Works:
- Graph Learning: Learns node embeddings and computes their similarity
- Mix-hop Propagation: Aggregates information from different hop distances
- Dilated Inception: Captures multi-scale temporal patterns
- Joint Learning: Graph structure and predictions are learned together
MTGNN Architecture:
- Node Embeddings: Each node gets a learnable embedding vector
- Adaptive Adjacency: A = softmax(E1 * E2^T) computes learned graph
- Mix-hop Propagation: Combines 1-hop, 2-hop, … K-hop neighbors
- Dilated Inception: Parallel dilated convolutions for temporal patterns
Key Benefits:
- No need to predefine graph structure
- Discovers hidden variable relationships
- Handles multivariate series with complex dependencies
- Combines best of GCN and TCN architectures
How It Works
MTGNN is a graph neural network that simultaneously learns the graph structure and performs spatio-temporal forecasting, without requiring a predefined adjacency matrix.
Reference: Wu et al., “Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks”, KDD 2020.
https://arxiv.org/abs/2005.11650
Constructors
| Constructor | Summary |
|---|---|
MTGNNOptions() |
Initializes a new instance of the MTGNNOptions class with default values. |
MTGNNOptions(MTGNNOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UsePredefinedGraph { get; set; } |
Gets or sets whether to use a predefined adjacency matrix alongside learned graph. |
bool UseSubgraphSampling { get; set; } |
Gets or sets whether to use subgraph sampling during training. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate for regularization. |
int DilationFactor { get; set; } |
Gets or sets the dilation factor for temporal convolutions. |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon. |
int HiddenDimension { get; set; } |
Gets or sets the hidden dimension for the model. |
int MixHopDepth { get; set; } |
Gets or sets the depth of mix-hop propagation. |
int NodeEmbeddingDim { get; set; } |
Gets or sets the dimension of node embeddings for graph learning. |
int NumLayers { get; set; } |
Gets or sets the number of layers in the model. |
int NumNodes { get; set; } |
Gets or sets the number of nodes (variables/time series). |
int NumSamples { get; set; } |
Gets or sets the number of samples for uncertainty estimation. |
int SequenceLength { get; set; } |
Gets or sets the sequence length (input time steps). |
int SubgraphSize { get; set; } |
Gets or sets the size of sampled subgraphs. |
int TemporalKernelSize { get; set; } |
Gets or sets the kernel size for temporal convolutions. |