LCE Prediction Object

Prediction object for TaskLCE. Carries the per-row true and predicted surrogate performance and, depending on the producing learner's predict type, a distributional payload describing the predictive of the performance curve f(b):

• response (numeric()): predictive median on the natural scale. For the sample-based learners this is the mean of the draws on the lce_link scale, back-transformed – which is the median whenever the link-scale distribution is symmetric (e.g. Gaussian). Always present.

• se (numeric()): total predictive standard deviation on the link scale (epistemic + aleatoric), the quantity to score against the realised y_b.

• se_epistemic (numeric()): standard deviation of the mean f(b) on the link scale (epistemic only), the quantity for expected-crossing decisions.

• quantiles (matrix()): predictive quantiles, one row per observation, one column per probability (carried in the "probs" attribute).

• samples (matrix()): predictive draws, one row per observation, one column per draw. Columns are joint sample paths for the sample-based learners.

• target_reached (matrix()): probability that the metric has reached a target, one row per observation, one column per target (carried in the "target" attribute).

The link scale on which se / se_epistemic live is a property of the TaskLCE (its link); the prediction itself stays a few plain numeric / matrix columns.

Creates a new PredictionLCE.

Arguments

task

(TaskLCE)
Task used to derive row ids and truth.

row_ids

(integer())
Row ids of the predictions.

truth

(numeric())
True surrogate performances.

response

(numeric())
Predicted surrogate performances (natural-scale predictive median).

se

(numeric())
Total predictive standard error on the link scale.

se_epistemic

(numeric())
Epistemic standard error of the mean on the link scale.

quantiles

(matrix())
Predicted quantiles (rows = observations, columns = probabilities). The probabilities must be stored in the "probs" attribute.

samples

(matrix())
Predictive draws (rows = observations, columns = draws).

target_reached

(matrix())
Reach probabilities (rows = observations, columns = targets). The targets must be stored in the "target" attribute.

weights

(numeric())
Optional measure weights.

check

(logical(1))
Whether to validate the inputs.

Fields

response

(numeric())
Predicted surrogate performance for each row.

se

(numeric())
Total predictive standard error (link scale), or NA vector when absent.

se_epistemic

(numeric())
Epistemic standard error of the mean (link scale), or NA when absent.

quantiles

(matrix())
Matrix of predicted quantiles (rows = observations, columns ascending).

samples

(matrix())
Matrix of predictive draws (rows = observations, columns = draws).

target_reached

(matrix())
Matrix of reach probabilities (rows = observations, columns = targets).