Do we really need `distr` prediction type in `PredictionRegr`?

[bblodfon]

I think this predict type was created because in mlr3proba we have a pipeline (and associated PipeOp and plotting function) to combine response and se regression predict types into some kind of distribution form (distr). I don't think we really need this, unless someone makes a more proper/dedicated probabilistic regression/distribution package.

See example of what the pipeline in mlr3proba does (note that the printing for the distr is not optimal, in mlr3proba we wrap this distr6 objects in a list in the data.table shown below):

library(mlr3proba)
#> Loading required package: mlr3
library(mlr3pipelines)
set.seed(1)
task = tsk("boston_housing")

# Option 1: Use a learner that can predict se
learn = lrn("regr.featureless", predict_type = "se")
pred = learn$train(task)$predict(task)
poc = po("compose_probregr") # SE => DISTR
poc$train(list(NULL, NULL))
#> $output
#> NULL
poc$predict(list(pred, pred))[[1]]
#> 
#> ── <PredictionRegr> for 506 observations: ──────────────────────────────────────
#>  row_ids truth response       se
#>        1  24.0 22.52885 9.182176
#>        2  21.6 22.52885 9.182176
#>        3  34.7 22.52885 9.182176
#>      ---   ---      ---      ---
#>      504  23.9 22.52885 9.182176
#>      505  22.0 22.52885 9.182176
#>      506  19.0 22.52885 9.182176
#>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          distr
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            ---
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list

# Option 2: Use two learners, one for response and the other for se
learn_response = lrn("regr.rpart")
learn_se = lrn("regr.featureless", predict_type = "se")
pred_response = learn_response$train(task)$predict(task)
pred_se = learn_se$train(task)$predict(task)
poc = po("compose_probregr") # (SE, RESPONSE) => DISTR
poc$train(list(NULL, NULL))
#> $output
#> NULL
poc$predict(list(pred_response, pred_se))[[1]]
#> 
#> ── <PredictionRegr> for 506 observations: ──────────────────────────────────────
#>  row_ids truth response       se
#>        1  24.0 24.53538 9.182176
#>        2  21.6 26.91481 9.182176
#>        3  34.7 32.96875 9.182176
#>      ---   ---      ---      ---
#>      504  23.9 24.53538 9.182176
#>      505  22.0 24.53538 9.182176
#>      506  19.0 20.96400 9.182176
#>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                          distr
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            ---
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list
#>  <VectorDistribution>\n  Inherits from: <DistributionWrapper>\n  Public:\n    alias: \n    cdf: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    cf: function (t, ...) \n    clone: function (deep = FALSE) \n    confidence: function (alpha = 0.95, sides = "both", median = FALSE) \n    correlation: function () \n    decorators: active binding\n    description: Vector of 506 Uniforms\n    distlist: active binding\n    dmax: active binding\n    dmin: active binding\n    entropy: function (base = 2, ...) \n    getParameterValue: function (id, ...) \n    ids: active binding\n    inf: active binding\n    initialize: function (distlist = NULL, distribution = NULL, params = NULL, \n    iqr: function () \n    kurtosis: function (excess = TRUE, ...) \n    kurtosisType: active binding\n    liesInSupport: function (x, all = TRUE, bound = FALSE) \n    liesInType: function (x, all = TRUE, bound = FALSE) \n    mean: function (...) \n    median: function () \n    mgf: function (t, ...) \n    mode: function (which = "all") \n    modelTable: active binding\n    name: Vector: 506 Uniforms\n    parameters: function (id = NULL) \n    pdf: function (..., log = FALSE, simplify = TRUE, data = NULL) \n    pgf: function (z, ...) \n    prec: function () \n    print: function (n = 2, ...) \n    properties: active binding\n    quantile: function (..., lower.tail = TRUE, log.p = FALSE, simplify = TRUE, \n    rand: function (n, simplify = TRUE) \n    setParameterValue: function (..., lst = list(...), error = "warn", resolveConflicts = FALSE) \n    short_name: Vec506Unif\n    skewness: function (...) \n    skewnessType: active binding\n    stdev: function () \n    strprint: function (n = 10) \n    summary: function (full = TRUE, ...) \n    sup: active binding\n    support: active binding\n    symmetry: active binding\n    traits: active binding\n    type: active binding\n    valueSupport: active binding\n    variance: function (...) \n    variateForm: active binding\n    workingSupport: function () \n    wrappedModels: function (model = NULL) \n  Private:\n    .cdf: function (x1, lower.tail, log.p) \n    .decorators: NULL\n    .distlist: FALSE\n    .isCdf: 1\n    .isPdf: 1\n    .isQuantile: 1\n    .isRand: 1\n    .log: TRUE\n    .modelTable: data.table, data.frame\n    .outerParameters: NULL\n    .parameters: ParameterSet, R6\n    .pdf: function (x1, log) \n    .properties: list\n    .quantile: function (x1, lower.tail, log.p) \n    .rand: function (x1) \n    .sharedparams: list\n    .traits: list\n    .trials: \n    .univariate: TRUE\n    .updateDecorators: function (decs) \n    .variates: 1\n    .wrappedModels: list

^{Created on 2025-06-26 with reprex v2.1.1}

[wmay]

Oh hey, I've been doing a lot of distributional regression with mlr3, and I'm planning to make an extension package for it.

Distributional regression is a big deal in weather forecasting (along with values that are downstream of weather, such as flooding and electrical loads). So it's pretty awesome that mlr3 supports it in the base package.

I guess it doesn't necessarily have to be in the base regression class for me to do this. But it's pretty neat and convenient that it does.

Also, at least one of the standard regression methods (ranger) can do distributional regression, though it's not implemented in the mlr3 package yet.

[bblodfon]

Hi Will (@wmay),

Thanks for your initiative! We will discuss this in the mlr3 workshop in two weeks. It seems keeping the distr prediction type and having a new mlr3-package that deals with distributional regression efficiently would be the way to go (using distr6 in mlr3proba was a real bottleneck, I guess distributional would be a better option here), lets see!

[wmay]

What is the workshop? I couldn't find any info about it.

I've been ok with distr6 so far, though I'm working with smaller datasets. I was actually just looking at the topmodels package from @zeileis for evaluating probabilistic forecasts, and saw he's using distributions3. But I don't think it would be hard for me to switch regardless.

[zeileis]

Thanks for linking me here. When we started developing topmodels (which I hope to finally push to CRAN this summer) we wanted something that was vectorized, could be easily stored in columns of a data frame, and does not add to many dependencies. And at that time distr and distr6 were not easily vectorized and distributional comes with many more dependencies. That's why we went with distributions3.

We also started talking to some package maintainers and could convince some of them to become topmodels-ready (especially the gamlss family). I would be happy to talk about adding methods for the mlr3 family.

[berndbischl]

hi guys. @wmay @zeileis

first, thx for geeting in contact with us.

1. the mlr3 workshop (currently happening) is more an internal "sprint" we do once per year, with the core team. but we are happy to collab with you in any form convenient.

2. currently, the situation with predict_type = "distr" is the following:
   a) distr6 is not on CRAN anymore -- and i dont think Raphael works too much on it anymore.
   b) no learner in mlr3 supports it
   c) we would really like to support Johns @bblodfon endeavours with survival -- but he might also go a different route.

3. To the point: From that perpective: We would rather remove the EXACT distr6 "distr" predict_type. BUT:
   We would be very happy to support (robust, in the sense of data type and software and engagement) distributional regression in mlr3.

---> Can we simply have a call about this?

[zeileis]

Cool, I'm in!

I'll be on vacation Aug 16-24 but the rest of Aug/Sep is still rather flexible. (In that time I might _possibly_ even make it to Munich in case we would want to hack a bit.) In Oct the new semester will swamp everything...

[wmay]

I'd be happy to have a call about it. My schedule is very flexible (though I guess my US time zone is a little off from both of yours)

[berndbischl]

@zeileis sorry for being super slow. i was a bit sick also and even missed the mlr3 workshop :(
spinal disc again.
shall we simply get on zoom now to discuss plans?
@wmay you are cordially invited, too

[zeileis]

Bernd @berndbischl thanks for the follow-up and sorry to hear about the spinal disc problems. I hope you're doing/getting better!

Zoom is fine, of course. I'm mostly available for the rest of the week (Wed to Fri), except Fri 10:00-13:00 CEST.

[wmay]

And I'm also still available!

[bblodfon]

Hi all, I have sent an email to arrange the call. Thanks for the interest! (let me know if you don't get it)

[bblodfon]

Dear all, short summary of the meeting:

• Achim showed us a demo of distributions3 (lightweight, vectorized, data.frame- based) + the use of prodist() function for distributional predictions and how the topmodels package scores these and produces various interesting plots.
• We will discuss internally how to best support this in mlr3 and have a follow-up meeting.

The above ofc implies that future mlr3 releases will become incompatible with the distr mlr3proba probabilistic regression prediction type as shown in the inital example code, but that's fine, as mlr3proba will be archived somewhere in 2026 and distr6 is not on CRAN.