DiffusionTSOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for DiffusionTS (Interpretable Diffusion for Time Series).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → DiffusionTSOptions<T>
For Beginners
DiffusionTS focuses on making diffusion models more interpretable by decomposing time series into understandable components:
The Key Insight: Time series often have clear structure (trends, seasonality) that gets lost in “black box” models. DiffusionTS preserves this structure by generating each component separately and combining them.
How DiffusionTS Works:
- Decomposition: Split time series into trend, seasonal, and residual
- Component Diffusion: Generate each component with specialized networks
- Reconstruction: Combine components to form final forecast
- Interpretation: Each component has clear meaning
DiffusionTS Architecture:
- Trend Network: Captures long-term movements (slow, smooth)
- Seasonal Network: Captures periodic patterns (daily, weekly, yearly)
- Residual Network: Captures irregular fluctuations
- Fusion Module: Combines components coherently
Key Benefits:
- Interpretable decomposition of forecasts
- Can enforce structural constraints (smooth trends, periodic seasons)
- Better uncertainty quantification per component
- Enables “what-if” analysis by modifying components
How It Works
DiffusionTS is an interpretable diffusion model for time series that uses seasonal-trend decomposition to generate forecasts with clear interpretable components.
Reference: Yuan and Qiu, “Diffusion-TS: Interpretable Diffusion for General Time Series Generation”, 2024.
https://arxiv.org/abs/2403.01742
Constructors
| Constructor | Summary |
|---|---|
DiffusionTSOptions() |
Creates options with the paper’s defaults. |
DiffusionTSOptions(DiffusionTSOptions<T> other) |
Copies every setting of other. |
Properties
| Property | Summary |
|---|---|
BetaSchedule BetaSchedule { get; set; } |
Gets or sets the noise schedule. |
DiffusionReconstructionLoss ReconstructionLoss { get; set; } |
Gets or sets the distance between the predicted and the true clean window. |
bool UseFourierLoss { get; set; } |
Gets or sets whether the Fourier-based loss term is added. |
double BetaEnd { get; set; } |
Gets or sets the last noise variance, used by the linear and scaled-linear schedules. |
double BetaStart { get; set; } |
Gets or sets the first noise variance, used by the linear and scaled-linear schedules. |
double DropoutRate { get; set; } |
Gets or sets the residual and embedding dropout. |
double? DenoisedClip { get; set; } |
Gets or sets the bound, in standard deviations of the context, that each predicted clean window is clamped to during sampling; null disables it. |
double? FourierLossWeight { get; set; } |
Gets or sets the weight of the Fourier loss term; null uses sqrt(window length) / 5. |
int ForecastHorizon { get; set; } |
Gets or sets the number of future steps to forecast. |
int FourierTopKFactor { get; set; } |
Gets or sets the factor of the seasonal block’s frequency count: it keeps int(factor x ln K) of the K retained frequencies. |
int HiddenDimension { get; set; } |
Gets or sets the transformer width d_model. |
int MlpHiddenTimes { get; set; } |
Gets or sets the MLP expansion of each block. |
int NumDecoderLayers { get; set; } |
Gets or sets the number of decoder blocks, each with its own trend and seasonal output. |
int NumDiffusionSteps { get; set; } |
Gets or sets the number of diffusion steps T. |
int NumEncoderLayers { get; set; } |
Gets or sets the number of encoder blocks. |
int NumHeads { get; set; } |
Gets or sets the number of attention heads. |
int NumSamples { get; set; } |
Gets or sets how many windows a forecast generates. |
int SequenceLength { get; set; } |
Gets or sets the number of past steps a forecast is conditioned on. |