API Reference¶

The public API is exported from the top-level dicex package; the exact solutions, the scaler and the optimizer live in their submodules.

Explainer¶

Class	Description
`Dicex`	High-level explainer: fits the scaler and returns the recommended direction for a record.
`ExplanationResult`	The direction, its robust value, the risk level and the details of the search.

Distributions of the random step length \(T\) and of the additive noise \(\varepsilon\) in the execution of a direction.

Class	Description
`GaussianPerturbation`	Gaussian step length (i.i.d., per-feature, or full covariance) plus additive noise.
`UniformPerturbation`	Step length with i.i.d. uniform coordinates on `[mu - delta, mu + delta]`, plus additive noise.
`CustomPerturbation`	Step length given by a user-defined sampling function.

Function	Description
`linear_directional_functional`, `linear_optimal_direction`	Closed-form objective and optimal direction for linear models under Gaussian perturbations.
`logistic_directional_functional`, `logistic_optimal_direction`	The same for logistic models.

Class	Description
`DicexScaler`	Standardization used by `Dicex`; maps directions back to the original units.
`VmfVns`	The vMF-VNS optimizer; its options can be passed to `Dicex` as keyword arguments.

Module	Description
`dicex.exceptions`	Errors and warnings raised by the library.