wfsets_penguins01

statlearning

tidymodels

num

wfsets

Published

May 17, 2023

Aufgabe

Berechnen Sie die Vorhersagegüte (RMSE) für folgende Lernalgorithmen:

lineares Modell
knn (neighbors: tune)

Modellgleichung: body_mass_g ~ bill_length_mm, data = d_train.

Nutzen Sie minimale Vorverarbeitung.

Lösung

Setup

library(tidymodels)
data(penguins, package = "palmerpenguins")

Daten

d <-
  penguins %>% 
  drop_na()

d_split <- initial_split(d)
d_train <- training(d_split)
d_test <- testing(d_split)

Modelle

Lineares Modell:

mod_lin <- linear_reg()

mod_knn <- nearest_neighbor(mode = "regression",
                                  neighbors = tune())

Rezepte

rec_basic <- recipe(body_mass_g ~ bill_length_mm, data = d_train) %>% 
         step_normalize(all_predictors())

Resampling

rsmpls <- vfold_cv(d_train)

Workflow Set

wf_set <-
  workflow_set(
    preproc = list(rec_simple = rec_basic),
    models = list(mod_lm = mod_lin,
                  mod_nn = mod_knn)
  )

Fitten

wf_fit <-
  wf_set %>% 
  workflow_map(resamples = rsmpls)

Check:

wf_fit %>% pluck("result")

Bester Kandidat

autoplot(wf_fit)

autoplot(wf_fit, select_best = TRUE)

rank_results(wf_fit, rank_metric = "rmse") %>% 
  filter(.metric == "rmse")

Am besten war das lineare Modell, aber schauen wir uns auch mal das knn-Modell an, v.a. um zu wissen, wie man den besten Tuningparameter-Wert sieht:

extract_workflow_set_result(wf_fit, "rec_simple_mod_nn") %>% 
  select_best()

Last Fit

best_wf <-
  wf_fit %>% 
  extract_workflow("rec_simple_mod_lm")

Finalisieren müssen wir diesen Workflow nicht, da er keine Tuningparameter hatte.

fit_final <-
  best_wf %>% 
  last_fit(d_split)

Modellgüte im Test-Set

collect_metrics(fit_final)

Categories:

R
statlearning
tidymodels
num

--- exname: wfsets_penguins01 extype: num exsolution: r fmt(sol) exshuffle: no expoints: 1 categories: - R - statlearning - tidymodels - num - wfsets date: '2023-05-17' slug: wfsets_penguins01 title: wfsets_penguins01 execute: output: false --- ```{r global-knitr-options, include=FALSE} knitr::opts_chunk$set(fig.pos = 'H', fig.asp = 0.618, fig.width = 4, fig.cap = "", fig.path = "", cache = TRUE, echo = TRUE, message = FALSE, fig.show = "hold") library(tidyverse) library(exams) ``` # Aufgabe Berechnen Sie die Vorhersagegüte (RMSE) für folgende Lernalgorithmen: - lineares Modell - knn (neighbors: tune) Modellgleichung: `body_mass_g ~ bill_length_mm, data = d_train`. Nutzen Sie minimale Vorverarbeitung. # Lösung ## Setup ```{r} library(tidymodels) data(penguins, package = "palmerpenguins") ``` ## Daten ```{r} d <- penguins %>% drop_na() ``` ```{r} d_split <- initial_split(d) d_train <- training(d_split) d_test <- testing(d_split) ``` ## Modelle Lineares Modell: ```{r} mod_lin <- linear_reg() mod_knn <- nearest_neighbor(mode = "regression", neighbors = tune()) ``` ## Rezepte ```{r} rec_basic <- recipe(body_mass_g ~ bill_length_mm, data = d_train) %>% step_normalize(all_predictors()) ``` ## Resampling ```{r} rsmpls <- vfold_cv(d_train) ``` ## Workflow Set ```{r} wf_set <- workflow_set( preproc = list(rec_simple = rec_basic), models = list(mod_lm = mod_lin, mod_nn = mod_knn) ) ``` ## Fitten ```{r} wf_fit <- wf_set %>% workflow_map(resamples = rsmpls) ``` Check: ```{r} wf_fit %>% pluck("result") ``` ## Bester Kandidat ```{r} autoplot(wf_fit) ``` ```{r} autoplot(wf_fit, select_best = TRUE) ``` ```{r} rank_results(wf_fit, rank_metric = "rmse") %>% filter(.metric == "rmse") ``` Am besten war das lineare Modell, aber schauen wir uns auch mal das knn-Modell an, v.a. um zu wissen, wie man den besten Tuningparameter-Wert sieht: ```{r} extract_workflow_set_result(wf_fit, "rec_simple_mod_nn") %>% select_best() ``` ## Last Fit ```{r} best_wf <- wf_fit %>% extract_workflow("rec_simple_mod_lm") ``` Finalisieren müssen wir diesen Workflow nicht, da er keine Tuningparameter hatte. ```{r} fit_final <- best_wf %>% last_fit(d_split) ``` ## Modellgüte im Test-Set ```{r} collect_metrics(fit_final) ``` --- Categories: - R - statlearning - tidymodels - num