CCDMOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for CCDM, the Channel-aware Contrastive Conditional Diffusion model for multivariate probabilistic time series forecasting (Li, Chen and Xiong, 2024).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → CCDMOptions<T>
For Beginners
CCDM forecasts by starting from random noise and removing it step by step, guided by the history it is given. Because every run starts from different noise, it draws many possible futures; their median is the forecast and their spread is the uncertainty.
How It Works
CCDM is a DDPM forecaster. Its denoiser embeds every variable’s past window and noisy future
independently (channel-independent dense modules), mixes the variables with channel-wise
diffusion transformers conditioned on the diffusion step through adaptive layer norm, and is
trained to predict the injected noise. An optional denoising-based InfoNCE term contrasts the
true future against shuffled and rescaled ones. Defaults follow the authors’ reference
configuration (github.com/LSY-Cython/CCDM, run_I48_O96.py) unless noted.
The inherited NumFeatures is the number of variables
(D) forecast jointly. Each variable is one token of the channel-wise diffusion transformers, so
attention runs ACROSS variables; inputs are laid out [context, variables] (optionally with a
leading batch) and forecasts [horizon, variables]. With one variable the attention reduces to a
single token.
Constructors
| Constructor | Summary |
|---|---|
CCDMOptions() |
Initializes a new instance with default values. |
CCDMOptions(CCDMOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
BetaSchedule BetaSchedule { get; set; } |
Gets or sets the shape of the noise schedule. |
double AttentionDropoutRate { get; set; } |
Gets or sets the dropout rate on the attention weights. |
double BetaEnd { get; set; } |
Gets or sets the last beta of the noise schedule. |
double BetaStart { get; set; } |
Gets or sets the first beta of the noise schedule. |
double ContrastiveTemperature { get; set; } |
Gets or sets the InfoNCE temperature tau. |
double ContrastiveWeight { get; set; } |
Gets or sets the weight lambda of the denoising-based temporal contrastive loss. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate inside the residual MLP blocks of the dense modules. |
double LearningRate { get; set; } |
Learning rate for the default Adam optimizer. |
double MlpRatio { get; set; } |
Gets or sets the hidden-width multiplier of the transformer MLP. |
int ContextLength { get; set; } |
Gets or sets the number of historical time steps used as input context (L). |
int DiffusionSteps { get; set; } |
Gets or sets the number of diffusion steps (K). |
int EmbeddingLayers { get; set; } |
Gets or sets the number of residual MLP blocks in each channel-independent dense module (n_emb; n_enc = n_dec in the paper). |
int ForecastHorizon { get; set; } |
Gets or sets the number of future time steps to forecast (H). |
int HiddenDimension { get; set; } |
Gets or sets the embedding width of the past window, the noisy future and the diffusion step (e_hid; the transformers run at twice this width). |
int NumHeads { get; set; } |
Gets or sets the number of attention heads. |
int NumLayers { get; set; } |
Gets or sets the number of channel-wise diffusion transformer blocks (n_att). |
int NumNegatives { get; set; } |
Gets or sets the number of negatives per augmentation (patch shuffle and magnitude scaling). |
int NumSamples { get; set; } |
Gets or sets the number of sample paths drawn per forecast (S). |
int TrainingBatchSize { get; set; } |
Number of (diffusion step, noise) draws averaged into a single training step. |