rtemis (version 1.4.3)

explain: Explain a prediction

Description

Per-case explanation of a fitted model: why it predicted what it did, for each case in newdata. Where get_varimp() ranks features across a whole model, this attributes one prediction -- a feature can rank third overall and be the entire reason for one case.

Usage

explain(x, newdata, background = NULL, config = NULL, verbosity = 1L, ...)

Value

SHAP object, or a named list of them for a SupervisedRes with type = "all".

Arguments

x

Supervised or SupervisedRes object.

newdata

tabular data: Cases to explain. Predictors only, in training order, as stats::predict() requires.

background

Optional tabular data: Cases the contributions are measured against, in the same shape as newdata. Required by every estimator that needs E[x] or a reference set, which is most of them; those abort with a message naming it when it is missing.

config

Optional ExplanationConfig object: Built by setup_SHAP. NULL uses setup_SHAP(), which resolves the estimator per algorithm.

verbosity

Integer: Verbosity level.

...

Additional arguments passed to methods, such as type for a SupervisedRes.

Author

EDG

Details

background is what the contributions are measured against, and most estimators cannot work without one. A contribution says how far a feature moved this case's prediction away from E[f(x)], and that expectation is a property of the background data, and the model does not store the data it was trained on. Pass the training features, or a representative sample of them. The exceptions are the estimators that take their baseline from the model itself, such as the LightGBM family's; those ignore it, and everything else aborts without it.

Two explanations of one model against different backgrounds are not comparable. The result carries a fingerprint of the background it used.

config chooses the method, built by setup_SHAP: setup_SHAP(perturbation = "conditional") picks the value function, setup_SHAP(estimator = "kernel") overrides the estimator the algorithm would otherwise get. explanation_traits reports which estimator applies to which algorithm, whether it is exact, and why.

A SupervisedRes additionally takes type: "avg" (the default) averages the resamples' explanations, which decomposes the averaged prediction exactly, and "all" returns one explanation per resample.

The contributions plus the baseline reconstruct the prediction exactly, on the scale they were computed on. For classification that scale is the model's margin, not the probability stats::predict() returns: probability is a nonlinear transform of the margin, so contributions do not sum to it.

Read the result with shap_case for one case, shap_long across cases, shap_by_level for the levels of a categorical, and get_varimp() for mean(|phi|) per feature.

A SHAP value is not a causal effect. It attributes a prediction under the model's own logic; intervening on a feature in the world does not move the outcome by its attribution. Correlated features split credit between themselves, and how they split it depends on the value function -- so a small attribution does not show a feature to be unimportant, and this is not a feature-selection criterion.

Examples

x <- data.frame(age = rnorm(100), bmi = rnorm(100))
x[["y"]] <- x[["age"]] * 2 + rnorm(100, sd = 0.3)
features <- x[, c("age", "bmi")]
mod <- train(x, hyperparameters = setup_GLM(), verbosity = 0L)

# `background` is what the contributions are measured against.
contributions <- explain(
  mod,
  features[1:5, ],
  background = features,
  verbosity = 0L
)
contributions

# Contributions plus the baseline reconstruct the prediction, exactly:
rowSums(contributions@phi[["outcome"]]) + contributions@baseline[, 1L]
predict(mod, features[1:5, ], verbosity = 0L)

# One case, ordered the way a waterfall reads:
shap_case(contributions, features[1:5, ], case = 1L)

# Which features mattered across these cases:
get_varimp(contributions)

# `config` chooses the method; here, a subsampled background.
explain(
  mod,
  features[1:5, ],
  background = features,
  config = setup_SHAP(background_n = 50L),
  verbosity = 0L
)