DCRNNOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for DCRNN (Diffusion Convolutional Recurrent Neural Network).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → DCRNNOptions<T>
For Beginners
DCRNN was specifically designed for traffic forecasting by combining two powerful ideas:
The Key Insight: Traffic flow on road networks can be modeled as a diffusion process - like how congestion spreads through a network. DCRNN captures this with diffusion convolution while using an encoder-decoder architecture for multi-step prediction.
What Problems Does DCRNN Solve?
- Traffic speed/flow prediction on road networks
- Air quality forecasting across sensor networks
- Subway ridership prediction
- Any spatial-temporal forecasting where diffusion dynamics matter
How DCRNN Works:
- Diffusion Convolution: Models spatial dependencies as bidirectional random walks
- Diffusion GRU: Replaces matrix multiplications in GRU with diffusion convolution
- Encoder-Decoder: Encoder captures history, decoder generates predictions
- Scheduled Sampling: Gradually transitions from ground truth to predictions during training
DCRNN Architecture:
- Encoder: Stacked DCGRU layers that encode input sequence
- Decoder: Stacked DCGRU layers that generate output sequence
- Diffusion: D_O^(K) + D_I^(K) bidirectional diffusion matrices
- Output: Linear projection to forecast values
Key Benefits:
- Captures spatial dependencies through diffusion process (not just adjacency)
- Multi-step prediction through encoder-decoder architecture
- Scheduled sampling prevents exposure bias during training
- Bidirectional diffusion captures both upstream and downstream effects
How It Works
DCRNN combines diffusion convolution with sequence-to-sequence architecture for spatial-temporal forecasting on graph-structured data.
Reference: Li et al., “Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting”, ICLR 2018.
https://arxiv.org/abs/1707.01926
Constructors
| Constructor | Summary |
|---|---|
DCRNNOptions() |
Initializes a new instance of the DCRNNOptions class with default values. |
DCRNNOptions(DCRNNOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseScheduledSampling { get; set; } |
Gets or sets whether to use scheduled sampling during training. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate for regularization. |
double LearningRate { get; set; } |
Gets or sets the initial Adam learning rate. |
double MinTeacherForcingRatio { get; set; } |
Gets or sets the minimum teacher forcing ratio. |
int DiffusionSteps { get; set; } |
Gets or sets the number of diffusion steps (K). |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon. |
int HiddenDimension { get; set; } |
Gets or sets the hidden dimension for the DCGRU cells. |
int MaxDiffusionStep { get; set; } |
Gets or sets the maximum diffusion step power. |
int NumDecoderLayers { get; set; } |
Gets or sets the number of decoder DCGRU layers. |
int NumEncoderLayers { get; set; } |
Gets or sets the number of encoder DCGRU layers. |
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 ScheduledSamplingDecaySteps { get; set; } |
Gets or sets the number of training steps for scheduled sampling decay. |
int SequenceLength { get; set; } |
Gets or sets the sequence length (input time steps). |
string FilterType { get; set; } |
Gets or sets the type of filter for diffusion. |