Skip to content

Expose and document VGP variational parameters for fit diagnostics #43

Description

@bwengals

The VGP Opper-Archambeau variational parameters carry per-point meaning that could be exposed for diagnostics. alpha is the expected score (a representer weight and pseudo-residual showing each point's influence) and lambda is the expected per-point information (the effective precision). Their reading depends on the likelihood: under Poisson lambda is the expected rate, under Bernoulli it peaks at the decision boundary, and under Student-t it drops toward zero for discounted outliers. Adding a helper to extract alpha, lambda, and diag(S) from a fit aligned to the training inputs, with the per-likelihood interpretations documented and checked against the Opper and Archambeau (2009) paper, would be worth doing. A skill file describing how to read these parameters for fit and data diagnosis should be included as well.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions