Methods

Latents implements methods from a family of probabilistic generative models. Each method defines a state model (distribution over latents) and an observation model (how observations arise from latents).

Method family

	1 Group	2 Groups	N Groups
Static	pPCA → FA	pCCA	GFA
GP time series	GPFA	DLAG	mDLAG

• Rows share a state model (static or Gaussian process time series)

• Columns share observation model structure (number of observed groups)

• pPCA → FA indicates FA generalizes pPCA (anisotropic vs isotropic noise)

• GFA and mDLAG use automatic relevance determination (ARD) for automatic dimensionality selection

Tip

See Mathematical background for the shared generative framework underlying all methods.

Choosing a method

How many observation groups do you have?

One group → single-group methods (left column). Multiple groups with shared latent structure → multi-group methods (right columns).

Is your data static or time series?

Independent samples → static methods (top row). Temporal dependencies → time series methods (bottom row).

Do you need automatic dimensionality selection?

If yes, use Bayesian methods with ARD: GFA (static) or mDLAG (time series).

Tip

Will this method work for my data? The Benchmarks show how runtime scales with dataset size and how well methods recover ground truth across different sample counts and dimensionalities — useful for assessing feasibility given your data constraints.

Available methods

Note

Latents currently implements GFA (complete) and mDLAG (under development).

• Group factor analysis (GFA)
• Delayed latents across multiple groups (mDLAG)