DCRNN<T>

Models & Types · AiDotNet.Finance.Graph

DCRNN (Diffusion Convolutional Recurrent Neural Network) for spatial-temporal forecasting.

Inheritance: NeuralNetworkBase<T> → FinancialModelBase<T> → ForecastingModelBase<T> → DCRNN<T>

For Beginners

DCRNN was a breakthrough model for traffic prediction that introduced two key innovations:

The Key Insight: Traffic flow can be modeled as a diffusion process - congestion spreads through a road network similar to how heat diffuses through a material. DCRNN captures this with diffusion convolution while using an encoder-decoder architecture for multi-step forecasting.

What Makes DCRNN Special:

  1. Diffusion Convolution: Models spatial dependencies as bidirectional random walks on the graph
  2. DCGRU Cells: GRU cells where matrix multiplications are replaced with diffusion convolution
  3. Encoder-Decoder: Seq2seq architecture for multi-step prediction
  4. Scheduled Sampling: Gradually transitions from teacher forcing to autoregressive during training

Mathematical Foundation: Diffusion convolution: X_star = sum_k (theta_k * (D_O^(-1)*W)^k * X + theta’_k * (D_I^(-1)*W^T)^k * X) where D_O, D_I are out/in-degree matrices and W is the adjacency matrix.

Architecture:

  • Encoder: Stacked DCGRU layers process input sequence
  • Final encoder state becomes initial decoder state
  • Decoder: Stacked DCGRU layers generate output autoregressively
  • Output: Linear projection to forecast dimension

Key Benefits:

  • Captures bidirectional spatial dependencies (upstream and downstream effects)
  • Multi-step prediction without error accumulation from teacher forcing
  • Effective on large-scale traffic networks

How It Works

DCRNN combines diffusion convolution with sequence-to-sequence architecture for traffic forecasting on road networks and other spatial-temporal prediction tasks.

Reference: Li et al., “Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting”, ICLR 2018. https://arxiv.org/abs/1707.01926

Example

// Define architecture for spatial-temporal traffic forecasting (207 sensors, 12-step horizon)
var architecture = new NeuralNetworkArchitecture<double>(
    inputType: InputType.OneDimensional,
    taskType: NeuralNetworkTaskType.Regression,
    inputHeight: 12, inputWidth: 207, inputDepth: 2, outputSize: 12);
            
// Training mode: diffusion convolutional recurrent network with graph structure
var model = new DCRNN<double>(architecture);
            
// ONNX inference mode: load pre-trained DCRNN model
var onnxModel = new DCRNN<double>(architecture, "dcrnn_traffic.onnx");

Constructors

Constructor Summary
DCRNN(NeuralNetworkArchitecture<T> architecture, DCRNNOptions<T> options, double[] adjacencyMatrix, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Initializes a new instance of DCRNN in native training/inference mode.
DCRNN(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, DCRNNOptions<T> options, double[] adjacencyMatrix, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) Initializes a new instance of DCRNN in ONNX inference mode.

Properties

Property Summary
bool IsChannelIndependent { get; }
bool SupportsTraining { get; } Gets whether the model supports training.
bool UseNativeMode { get; }
int DiffusionSteps { get; } Gets the number of diffusion steps.
int ForecastHorizon { get; } Gets the forecast horizon.
int NumFeatures { get; }
int NumNodes { get; } Gets the number of nodes in the graph.
int NumSamples { get; } Gets the number of samples for uncertainty estimation.
int PatchSize { get; }
int PredictionHorizon { get; }
int SequenceLength { get; }
int Stride { get; }

Methods

Method Summary
Dictionary<string, T> Evaluate(Tensor<T> predictions, Tensor<T> actuals) Evaluates forecast quality against actual values.
Dictionary<string, T> GetFinancialMetrics() Gets financial-specific metrics.
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input)
ModelInputShapeConstraint GetInputShapeConstraint() The input DCRNN reads: [numNodes, sequenceLength, featuresPerNode].
ModelMetadata<T> GetModelMetadata() Gets metadata about the DCRNN model.
Tensor<T> ApplyInstanceNormalization(Tensor<T> input) Applies instance normalization (identity for DCRNN).
Tensor<T> AutoregressiveForecast(Tensor<T> input, int steps) Performs autoregressive forecasting for extended horizons.
Tensor<T> Forward(Tensor<T> input) Performs forward pass through all layers.
Tensor<T> ForwardForTraining(Tensor<T> input) Runs tape training through the same last-step readout and graph-diffusion path used by native forecasting.
ValueTuple<Tensor<T>, Tensor<T>, Tensor<T>> ForecastWithIntervals(Tensor<T> input, double confidenceLevel) Generates forecasts with prediction intervals.