{ "nbformat": 4, "nbformat_minor": 2, "metadata": { "colab": { "name": "lesson_11-R.ipynb", "provenance": [], "collapsed_sections": [], "toc_visible": true }, "kernelspec": { "name": "ir", "display_name": "R" }, "language_info": { "name": "R" }, "coopTranslator": { "original_hash": "6ea6a5171b1b99b7b5a55f7469c048d2", "translation_date": "2025-09-06T12:26:27+00:00", "source_file": "4-Classification/2-Classifiers-1/solution/R/lesson_11-R.ipynb", "language_code": "no" } }, "cells": [ { "cell_type": "markdown", "source": [], "metadata": { "id": "zs2woWv_HoE8" } }, { "cell_type": "markdown", "source": [ "## Klassifisering av matretter 1\n", "\n", "I denne leksjonen skal vi utforske ulike klassifiseringsmetoder for å *forutsi en nasjonal matrett basert på en gruppe ingredienser.* Samtidig skal vi lære mer om hvordan algoritmer kan brukes til klassifiseringsoppgaver.\n", "\n", "### [**Quiz før leksjonen**](https://gray-sand-07a10f403.1.azurestaticapps.net/quiz/21/)\n", "\n", "### **Forberedelse**\n", "\n", "Denne leksjonen bygger videre på vår [forrige leksjon](https://github.com/microsoft/ML-For-Beginners/blob/main/4-Classification/1-Introduction/solution/lesson_10-R.ipynb) der vi:\n", "\n", "- Ga en enkel introduksjon til klassifisering ved hjelp av et datasett om alle de fantastiske matrettene fra Asia og India 😋.\n", "\n", "- Utforsket noen [dplyr-kommandoer](https://dplyr.tidyverse.org/) for å klargjøre og rense dataene våre.\n", "\n", "- Lagde vakre visualiseringer med ggplot2.\n", "\n", "- Viste hvordan man håndterer ubalanserte data ved å forhåndsbehandle dem med [recipes](https://recipes.tidymodels.org/articles/Simple_Example.html).\n", "\n", "- Demonstrerte hvordan man bruker `prep` og `bake` for å bekrefte at oppskriften fungerer som den skal.\n", "\n", "#### **Forutsetninger**\n", "\n", "For denne leksjonen trenger vi følgende pakker for å rense, klargjøre og visualisere dataene våre:\n", "\n", "- `tidyverse`: [tidyverse](https://www.tidyverse.org/) er en [samling av R-pakker](https://www.tidyverse.org/packages) designet for å gjøre datavitenskap raskere, enklere og morsommere!\n", "\n", "- `tidymodels`: [tidymodels](https://www.tidymodels.org/) er et [rammeverk med pakker](https://www.tidymodels.org/packages/) for modellering og maskinlæring.\n", "\n", "- `themis`: [themis-pakken](https://themis.tidymodels.org/) gir ekstra oppskritt for å håndtere ubalanserte data.\n", "\n", "- `nnet`: [nnet-pakken](https://cran.r-project.org/web/packages/nnet/nnet.pdf) tilbyr funksjoner for å estimere feed-forward nevrale nettverk med ett skjult lag, samt for modeller for multinomial logistisk regresjon.\n", "\n", "Du kan installere dem slik:\n" ], "metadata": { "id": "iDFOb3ebHwQC" } }, { "cell_type": "markdown", "source": [ "`install.packages(c(\"tidyverse\", \"tidymodels\", \"DataExplorer\", \"here\"))`\n", "\n", "Alternativt sjekker skriptet nedenfor om du har de nødvendige pakkene for å fullføre denne modulen, og installerer dem for deg hvis de mangler.\n" ], "metadata": { "id": "4V85BGCjII7F" } }, { "cell_type": "code", "execution_count": 2, "source": [ "suppressWarnings(if (!require(\"pacman\"))install.packages(\"pacman\"))\r\n", "\r\n", "pacman::p_load(tidyverse, tidymodels, themis, here)" ], "outputs": [ { "output_type": "stream", "name": "stderr", "text": [ "Loading required package: pacman\n", "\n" ] } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/" }, "id": "an5NPyyKIKNR", "outputId": "834d5e74-f4b8-49f9-8ab5-4c52ff2d7bc8" } }, { "cell_type": "markdown", "source": [ "## 1. Del dataene inn i trenings- og testsett.\n", "\n", "Vi starter med å velge noen få steg fra vår forrige leksjon.\n", "\n", "### Fjern de vanligste ingrediensene som skaper forvirring mellom ulike kjøkken, ved hjelp av `dplyr::select()`.\n", "\n", "Alle elsker ris, hvitløk og ingefær!\n" ], "metadata": { "id": "0ax9GQLBINVv" } }, { "cell_type": "code", "execution_count": 3, "source": [ "# Load the original cuisines data\r\n", "df <- read_csv(file = \"https://raw.githubusercontent.com/microsoft/ML-For-Beginners/main/4-Classification/data/cuisines.csv\")\r\n", "\r\n", "# Drop id column, rice, garlic and ginger from our original data set\r\n", "df_select <- df %>% \r\n", " select(-c(1, rice, garlic, ginger)) %>%\r\n", " # Encode cuisine column as categorical\r\n", " mutate(cuisine = factor(cuisine))\r\n", "\r\n", "# Display new data set\r\n", "df_select %>% \r\n", " slice_head(n = 5)\r\n", "\r\n", "# Display distribution of cuisines\r\n", "df_select %>% \r\n", " count(cuisine) %>% \r\n", " arrange(desc(n))" ], "outputs": [ { "output_type": "stream", "name": "stderr", "text": [ "New names:\n", "* `` -> ...1\n", "\n", "\u001b[1m\u001b[1mRows: \u001b[1m\u001b[22m\u001b[34m\u001b[34m2448\u001b[34m\u001b[39m \u001b[1m\u001b[1mColumns: \u001b[1m\u001b[22m\u001b[34m\u001b[34m385\u001b[34m\u001b[39m\n", "\n", "\u001b[36m──\u001b[39m \u001b[1m\u001b[1mColumn specification\u001b[1m\u001b[22m \u001b[36m────────────────────────────────────────────────────────\u001b[39m\n", "\u001b[1mDelimiter:\u001b[22m \",\"\n", "\u001b[31mchr\u001b[39m (1): cuisine\n", "\u001b[32mdbl\u001b[39m (384): ...1, almond, angelica, anise, anise_seed, apple, apple_brandy, a...\n", "\n", "\n", "\u001b[36mℹ\u001b[39m Use \u001b[30m\u001b[47m\u001b[30m\u001b[47m`spec()`\u001b[47m\u001b[30m\u001b[49m\u001b[39m to retrieve the full column specification for this data.\n", "\u001b[36mℹ\u001b[39m Specify the column types or set \u001b[30m\u001b[47m\u001b[30m\u001b[47m`show_col_types = FALSE`\u001b[47m\u001b[30m\u001b[49m\u001b[39m to quiet this message.\n", "\n" ] }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine almond angelica anise anise_seed apple apple_brandy apricot armagnac\n", "1 indian 0 0 0 0 0 0 0 0 \n", "2 indian 1 0 0 0 0 0 0 0 \n", "3 indian 0 0 0 0 0 0 0 0 \n", "4 indian 0 0 0 0 0 0 0 0 \n", "5 indian 0 0 0 0 0 0 0 0 \n", " artemisia ⋯ whiskey white_bread white_wine whole_grain_wheat_flour wine wood\n", "1 0 ⋯ 0 0 0 0 0 0 \n", "2 0 ⋯ 0 0 0 0 0 0 \n", "3 0 ⋯ 0 0 0 0 0 0 \n", "4 0 ⋯ 0 0 0 0 0 0 \n", "5 0 ⋯ 0 0 0 0 0 0 \n", " yam yeast yogurt zucchini\n", "1 0 0 0 0 \n", "2 0 0 0 0 \n", "3 0 0 0 0 \n", "4 0 0 0 0 \n", "5 0 0 1 0 " ], "text/markdown": [ "\n", "A tibble: 5 × 381\n", "\n", "| cuisine <fct> | almond <dbl> | angelica <dbl> | anise <dbl> | anise_seed <dbl> | apple <dbl> | apple_brandy <dbl> | apricot <dbl> | armagnac <dbl> | artemisia <dbl> | ⋯ ⋯ | whiskey <dbl> | white_bread <dbl> | white_wine <dbl> | whole_grain_wheat_flour <dbl> | wine <dbl> | wood <dbl> | yam <dbl> | yeast <dbl> | yogurt <dbl> | zucchini <dbl> |\n", "|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 1 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| indian | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1 | 0 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 381\n", "\\begin{tabular}{lllllllllllllllllllll}\n", " cuisine & almond & angelica & anise & anise\\_seed & apple & apple\\_brandy & apricot & armagnac & artemisia & ⋯ & whiskey & white\\_bread & white\\_wine & whole\\_grain\\_wheat\\_flour & wine & wood & yam & yeast & yogurt & zucchini\\\\\n", " & & & & & & & & & & ⋯ & & & & & & & & & & \\\\\n", "\\hline\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t indian & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 381
cuisinealmondangelicaaniseanise_seedappleapple_brandyapricotarmagnacartemisia⋯whiskeywhite_breadwhite_winewhole_grain_wheat_flourwinewoodyamyeastyogurtzucchini
<fct><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>⋯<dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>
indian000000000⋯0000000000
indian100000000⋯0000000000
indian000000000⋯0000000000
indian000000000⋯0000000000
indian000000000⋯0000000010
\n" ] }, "metadata": {} }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine n \n", "1 korean 799\n", "2 indian 598\n", "3 chinese 442\n", "4 japanese 320\n", "5 thai 289" ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | n <int> |\n", "|---|---|\n", "| korean | 799 |\n", "| indian | 598 |\n", "| chinese | 442 |\n", "| japanese | 320 |\n", "| thai | 289 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & n\\\\\n", " & \\\\\n", "\\hline\n", "\t korean & 799\\\\\n", "\t indian & 598\\\\\n", "\t chinese & 442\\\\\n", "\t japanese & 320\\\\\n", "\t thai & 289\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisinen
<fct><int>
korean 799
indian 598
chinese 442
japanese320
thai 289
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 735 }, "id": "jhCrrH22IWVR", "outputId": "d444a85c-1d8b-485f-bc4f-8be2e8f8217c" } }, { "cell_type": "markdown", "source": [ "Perfekt! Nå er det på tide å dele dataene slik at 70 % av dataene går til trening og 30 % til testing. Vi vil også bruke en `stratifisering`-teknikk når vi deler dataene for å `bevare andelen av hver matkultur` i trenings- og valideringsdatasett.\n", "\n", "[rsample](https://rsample.tidymodels.org/), en pakke i Tidymodels, gir infrastruktur for effektiv deling og resampling av data:\n" ], "metadata": { "id": "AYTjVyajIdny" } }, { "cell_type": "code", "execution_count": 4, "source": [ "# Load the core Tidymodels packages into R session\r\n", "library(tidymodels)\r\n", "\r\n", "# Create split specification\r\n", "set.seed(2056)\r\n", "cuisines_split <- initial_split(data = df_select,\r\n", " strata = cuisine,\r\n", " prop = 0.7)\r\n", "\r\n", "# Extract the data in each split\r\n", "cuisines_train <- training(cuisines_split)\r\n", "cuisines_test <- testing(cuisines_split)\r\n", "\r\n", "# Print the number of cases in each split\r\n", "cat(\"Training cases: \", nrow(cuisines_train), \"\\n\",\r\n", " \"Test cases: \", nrow(cuisines_test), sep = \"\")\r\n", "\r\n", "# Display the first few rows of the training set\r\n", "cuisines_train %>% \r\n", " slice_head(n = 5)\r\n", "\r\n", "\r\n", "# Display distribution of cuisines in the training set\r\n", "cuisines_train %>% \r\n", " count(cuisine) %>% \r\n", " arrange(desc(n))" ], "outputs": [ { "output_type": "stream", "name": "stdout", "text": [ "Training cases: 1712\n", "Test cases: 736" ] }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine almond angelica anise anise_seed apple apple_brandy apricot armagnac\n", "1 chinese 0 0 0 0 0 0 0 0 \n", "2 chinese 0 0 0 0 0 0 0 0 \n", "3 chinese 0 0 0 0 0 0 0 0 \n", "4 chinese 0 0 0 0 0 0 0 0 \n", "5 chinese 0 0 0 0 0 0 0 0 \n", " artemisia ⋯ whiskey white_bread white_wine whole_grain_wheat_flour wine wood\n", "1 0 ⋯ 0 0 0 0 1 0 \n", "2 0 ⋯ 0 0 0 0 1 0 \n", "3 0 ⋯ 0 0 0 0 0 0 \n", "4 0 ⋯ 0 0 0 0 0 0 \n", "5 0 ⋯ 0 0 0 0 0 0 \n", " yam yeast yogurt zucchini\n", "1 0 0 0 0 \n", "2 0 0 0 0 \n", "3 0 0 0 0 \n", "4 0 0 0 0 \n", "5 0 0 0 0 " ], "text/markdown": [ "\n", "A tibble: 5 × 381\n", "\n", "| cuisine <fct> | almond <dbl> | angelica <dbl> | anise <dbl> | anise_seed <dbl> | apple <dbl> | apple_brandy <dbl> | apricot <dbl> | armagnac <dbl> | artemisia <dbl> | ⋯ ⋯ | whiskey <dbl> | white_bread <dbl> | white_wine <dbl> | whole_grain_wheat_flour <dbl> | wine <dbl> | wood <dbl> | yam <dbl> | yeast <dbl> | yogurt <dbl> | zucchini <dbl> |\n", "|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 1 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "| chinese | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | ⋯ | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 381\n", "\\begin{tabular}{lllllllllllllllllllll}\n", " cuisine & almond & angelica & anise & anise\\_seed & apple & apple\\_brandy & apricot & armagnac & artemisia & ⋯ & whiskey & white\\_bread & white\\_wine & whole\\_grain\\_wheat\\_flour & wine & wood & yam & yeast & yogurt & zucchini\\\\\n", " & & & & & & & & & & ⋯ & & & & & & & & & & \\\\\n", "\\hline\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\t chinese & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & ⋯ & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & 0\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 381
cuisinealmondangelicaaniseanise_seedappleapple_brandyapricotarmagnacartemisia⋯whiskeywhite_breadwhite_winewhole_grain_wheat_flourwinewoodyamyeastyogurtzucchini
<fct><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>⋯<dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl><dbl>
chinese000000000⋯0000100000
chinese000000000⋯0000100000
chinese000000000⋯0000000000
chinese000000000⋯0000000000
chinese000000000⋯0000000000
\n" ] }, "metadata": {} }, { "output_type": "display_data", "data": { "text/plain": [ " cuisine n \n", "1 korean 559\n", "2 indian 418\n", "3 chinese 309\n", "4 japanese 224\n", "5 thai 202" ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | n <int> |\n", "|---|---|\n", "| korean | 559 |\n", "| indian | 418 |\n", "| chinese | 309 |\n", "| japanese | 224 |\n", "| thai | 202 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & n\\\\\n", " & \\\\\n", "\\hline\n", "\t korean & 559\\\\\n", "\t indian & 418\\\\\n", "\t chinese & 309\\\\\n", "\t japanese & 224\\\\\n", "\t thai & 202\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisinen
<fct><int>
korean 559
indian 418
chinese 309
japanese224
thai 202
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 535 }, "id": "w5FWIkEiIjdN", "outputId": "2e195fd9-1a8f-4b91-9573-cce5582242df" } }, { "cell_type": "markdown", "source": [ "## 2. Håndtere ubalanserte data\n", "\n", "Som du kanskje har lagt merke til i det opprinnelige datasettet, så vel som i treningssettet vårt, er det en ganske ujevn fordeling i antall kjøkken. Koreanske kjøkken er *nesten* 3 ganger så mange som thailandske kjøkken. Ubalanserte data har ofte negative effekter på modellens ytelse. Mange modeller fungerer best når antall observasjoner er likt, og har derfor en tendens til å slite med ubalanserte data.\n", "\n", "Det finnes hovedsakelig to måter å håndtere ubalanserte datasett på:\n", "\n", "- legge til observasjoner i minoritetsklassen: `Over-sampling`, for eksempel ved å bruke en SMOTE-algoritme som syntetisk genererer nye eksempler av minoritetsklassen ved hjelp av nærmeste naboer til disse tilfellene.\n", "\n", "- fjerne observasjoner fra majoritetsklassen: `Under-sampling`\n", "\n", "I vår forrige leksjon demonstrerte vi hvordan man kan håndtere ubalanserte datasett ved hjelp av en `recipe`. En recipe kan betraktes som en oppskrift som beskriver hvilke steg som skal brukes på et datasett for å gjøre det klart for dataanalyse. I vårt tilfelle ønsker vi å ha en lik fordeling i antall kjøkken i vårt `treningssett`. La oss sette i gang!\n" ], "metadata": { "id": "daBi9qJNIwqW" } }, { "cell_type": "code", "execution_count": 5, "source": [ "# Load themis package for dealing with imbalanced data\r\n", "library(themis)\r\n", "\r\n", "# Create a recipe for preprocessing training data\r\n", "cuisines_recipe <- recipe(cuisine ~ ., data = cuisines_train) %>% \r\n", " step_smote(cuisine)\r\n", "\r\n", "# Print recipe\r\n", "cuisines_recipe" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "Data Recipe\n", "\n", "Inputs:\n", "\n", " role #variables\n", " outcome 1\n", " predictor 380\n", "\n", "Operations:\n", "\n", "SMOTE based on cuisine" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 200 }, "id": "Az6LFBGxI1X0", "outputId": "29d71d85-64b0-4e62-871e-bcd5398573b6" } }, { "cell_type": "markdown", "source": [ "Du kan selvfølgelig gå videre og bekrefte (ved å forberede og bake) at oppskriften fungerer som forventet - alle matetikettene har `559` observasjoner.\n", "\n", "Siden vi skal bruke denne oppskriften som en forprosessering for modellering, vil en `workflow()` ta seg av all forberedelse og baking for oss, så vi slipper å estimere oppskriften manuelt.\n", "\n", "Nå er vi klare til å trene en modell 👩‍💻👨‍💻!\n", "\n", "## 3. Velge din klassifikator\n", "\n", "

\n", " \n", "

Kunstverk av @allison_horst
\n" ], "metadata": { "id": "NBL3PqIWJBBB" } }, { "cell_type": "markdown", "source": [ "Nå må vi bestemme oss for hvilken algoritme vi skal bruke til oppgaven 🤔.\n", "\n", "I Tidymodels gir [`parsnip-pakken`](https://parsnip.tidymodels.org/index.html) en konsekvent grensesnitt for å jobbe med modeller på tvers av ulike motorer (pakker). Se dokumentasjonen for parsnip for å utforske [modelltyper og motorer](https://www.tidymodels.org/find/parsnip/#models) og deres tilhørende [modellargumenter](https://www.tidymodels.org/find/parsnip/#model-args). Utvalget kan virke ganske overveldende ved første øyekast. For eksempel inkluderer følgende metoder klassifiseringsteknikker:\n", "\n", "- C5.0 Regelbaserte klassifiseringsmodeller\n", "\n", "- Fleksible diskriminantmodeller\n", "\n", "- Lineære diskriminantmodeller\n", "\n", "- Regulariserte diskriminantmodeller\n", "\n", "- Logistiske regresjonsmodeller\n", "\n", "- Multinomiale regresjonsmodeller\n", "\n", "- Naive Bayes-modeller\n", "\n", "- Støttevektormaskiner\n", "\n", "- Nærmeste naboer\n", "\n", "- Beslutningstrær\n", "\n", "- Ensemble-metoder\n", "\n", "- Nevrale nettverk\n", "\n", "Listen fortsetter!\n", "\n", "### **Hvilken klassifikator skal vi velge?**\n", "\n", "Så, hvilken klassifikator bør du velge? Ofte er det en god idé å teste flere og se etter et godt resultat.\n", "\n", "> AutoML løser dette problemet elegant ved å kjøre disse sammenligningene i skyen, slik at du kan velge den beste algoritmen for dine data. Prøv det [her](https://docs.microsoft.com/learn/modules/automate-model-selection-with-azure-automl/?WT.mc_id=academic-77952-leestott)\n", "\n", "Valget av klassifikator avhenger også av problemet vårt. For eksempel, når utfallet kan kategoriseres i `mer enn to klasser`, som i vårt tilfelle, må du bruke en `multiklasse klassifikasjonsalgoritme` i stedet for `binær klassifikasjon.`\n", "\n", "### **En bedre tilnærming**\n", "\n", "En bedre tilnærming enn å gjette vilt er å følge ideene i dette nedlastbare [ML Cheat Sheet](https://docs.microsoft.com/azure/machine-learning/algorithm-cheat-sheet?WT.mc_id=academic-77952-leestott). Her oppdager vi at, for vårt multiklasseproblem, har vi noen valg:\n", "\n", "

\n", " \n", "

En del av Microsofts Algorithm Cheat Sheet, som viser alternativer for multiklasse klassifikasjon
\n" ], "metadata": { "id": "a6DLAZ3vJZ14" } }, { "cell_type": "markdown", "source": [ "### **Resonnement**\n", "\n", "La oss se om vi kan resonnere oss fram til ulike tilnærminger gitt de begrensningene vi har:\n", "\n", "- **Dype nevrale nettverk er for tunge**. Gitt vårt rene, men minimale datasett, og det faktum at vi kjører trening lokalt via notatbøker, er dype nevrale nettverk for tunge for denne oppgaven.\n", "\n", "- **Ingen to-klassers klassifikator**. Vi bruker ikke en to-klassers klassifikator, så det utelukker one-vs-all.\n", "\n", "- **Beslutningstre eller logistisk regresjon kan fungere**. Et beslutningstre kan fungere, eller multinomial regresjon/multiklass logistisk regresjon for multiklass-data.\n", "\n", "- **Multiklass Boosted Decision Trees løser et annet problem**. Multiklass Boosted Decision Trees er mest egnet for ikke-parametriske oppgaver, f.eks. oppgaver designet for å bygge rangeringer, så det er ikke nyttig for oss.\n", "\n", "Vanligvis, før man går i gang med mer komplekse maskinlæringsmodeller som ensemble-metoder, er det en god idé å bygge den enkleste mulige modellen for å få en idé om hva som skjer. Så for denne leksjonen starter vi med en `multinomial regresjonsmodell`.\n", "\n", "> Logistisk regresjon er en teknikk som brukes når utfallsvariabelen er kategorisk (eller nominell). For binær logistisk regresjon er antall utfallsvariabler to, mens antall utfallsvariabler for multinomial logistisk regresjon er mer enn to. Se [Avanserte regresjonsmetoder](https://bookdown.org/chua/ber642_advanced_regression/multinomial-logistic-regression.html) for mer informasjon.\n", "\n", "## 4. Tren og evaluer en Multinomial logistisk regresjonsmodell.\n", "\n", "I Tidymodels definerer `parsnip::multinom_reg()` en modell som bruker lineære prediktorer for å forutsi multiklass-data ved hjelp av multinomial fordeling. Se `?multinom_reg()` for de ulike måtene/motorene du kan bruke for å tilpasse denne modellen.\n", "\n", "For dette eksempelet skal vi tilpasse en Multinomial regresjonsmodell via standardmotoren [nnet](https://cran.r-project.org/web/packages/nnet/nnet.pdf).\n", "\n", "> Jeg valgte en verdi for `penalty` litt tilfeldig. Det finnes bedre måter å velge denne verdien på, nemlig ved å bruke `resampling` og `tuning` av modellen, som vi skal diskutere senere.\n", ">\n", "> Se [Tidymodels: Kom i gang](https://www.tidymodels.org/start/tuning/) hvis du vil lære mer om hvordan du tuner modellens hyperparametere.\n" ], "metadata": { "id": "gWMsVcbBJemu" } }, { "cell_type": "code", "execution_count": 6, "source": [ "# Create a multinomial regression model specification\r\n", "mr_spec <- multinom_reg(penalty = 1) %>% \r\n", " set_engine(\"nnet\", MaxNWts = 2086) %>% \r\n", " set_mode(\"classification\")\r\n", "\r\n", "# Print model specification\r\n", "mr_spec" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "Multinomial Regression Model Specification (classification)\n", "\n", "Main Arguments:\n", " penalty = 1\n", "\n", "Engine-Specific Arguments:\n", " MaxNWts = 2086\n", "\n", "Computational engine: nnet \n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 166 }, "id": "Wq_fcyQiJvfG", "outputId": "c30449c7-3864-4be7-f810-72a003743e2d" } }, { "cell_type": "markdown", "source": [ "Flott jobb 🥳! Nå som vi har en oppskrift og en modellspesifikasjon, må vi finne en måte å samle dem sammen i et objekt som først vil forhåndsbehandle dataene, deretter tilpasse modellen på de forhåndsbehandlede dataene, og også tillate potensielle etterbehandlingsaktiviteter. I Tidymodels kalles dette praktiske objektet en [`workflow`](https://workflows.tidymodels.org/) og holder praktisk dine modelleringskomponenter! Dette er det vi ville kalt *pipelines* i *Python*.\n", "\n", "Så la oss samle alt i en workflow!📦\n" ], "metadata": { "id": "NlSbzDfgJ0zh" } }, { "cell_type": "code", "execution_count": 7, "source": [ "# Bundle recipe and model specification\r\n", "mr_wf <- workflow() %>% \r\n", " add_recipe(cuisines_recipe) %>% \r\n", " add_model(mr_spec)\r\n", "\r\n", "# Print out workflow\r\n", "mr_wf" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "══ Workflow ════════════════════════════════════════════════════════════════════\n", "\u001b[3mPreprocessor:\u001b[23m Recipe\n", "\u001b[3mModel:\u001b[23m multinom_reg()\n", "\n", "── Preprocessor ────────────────────────────────────────────────────────────────\n", "1 Recipe Step\n", "\n", "• step_smote()\n", "\n", "── Model ───────────────────────────────────────────────────────────────────────\n", "Multinomial Regression Model Specification (classification)\n", "\n", "Main Arguments:\n", " penalty = 1\n", "\n", "Engine-Specific Arguments:\n", " MaxNWts = 2086\n", "\n", "Computational engine: nnet \n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 333 }, "id": "Sc1TfPA4Ke3_", "outputId": "82c70013-e431-4e7e-cef6-9fcf8aad4a6c" } }, { "cell_type": "markdown", "source": [ "Arbeidsflyter 👌👌! En **`workflow()`** kan tilpasses på omtrent samme måte som en modell kan. Så, tid for å trene en modell!\n" ], "metadata": { "id": "TNQ8i85aKf9L" } }, { "cell_type": "code", "execution_count": 8, "source": [ "# Train a multinomial regression model\n", "mr_fit <- fit(object = mr_wf, data = cuisines_train)\n", "\n", "mr_fit" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "══ Workflow [trained] ══════════════════════════════════════════════════════════\n", "\u001b[3mPreprocessor:\u001b[23m Recipe\n", "\u001b[3mModel:\u001b[23m multinom_reg()\n", "\n", "── Preprocessor ────────────────────────────────────────────────────────────────\n", "1 Recipe Step\n", "\n", "• step_smote()\n", "\n", "── Model ───────────────────────────────────────────────────────────────────────\n", "Call:\n", "nnet::multinom(formula = ..y ~ ., data = data, decay = ~1, MaxNWts = ~2086, \n", " trace = FALSE)\n", "\n", "Coefficients:\n", " (Intercept) almond angelica anise anise_seed apple\n", "indian 0.19723325 0.2409661 0 -5.004955e-05 -0.1657635 -0.05769734\n", "japanese 0.13961959 -0.6262400 0 -1.169155e-04 -0.4893596 -0.08585717\n", "korean 0.22377347 -0.1833485 0 -5.560395e-05 -0.2489401 -0.15657804\n", "thai -0.04336577 -0.6106258 0 4.903828e-04 -0.5782866 0.63451105\n", " apple_brandy apricot armagnac artemisia artichoke asparagus\n", "indian 0 0.37042636 0 -0.09122797 0 -0.27181970\n", "japanese 0 0.28895643 0 -0.12651100 0 0.14054037\n", "korean 0 -0.07981259 0 0.55756709 0 -0.66979948\n", "thai 0 -0.33160904 0 -0.10725182 0 -0.02602152\n", " avocado bacon baked_potato balm banana barley\n", "indian -0.46624197 0.16008055 0 0 -0.2838796 0.2230625\n", "japanese 0.90341344 0.02932727 0 0 -0.4142787 2.0953906\n", "korean -0.06925382 -0.35804134 0 0 -0.2686963 -0.7233404\n", "thai -0.21473955 -0.75594439 0 0 0.6784880 -0.4363320\n", " bartlett_pear basil bay bean beech\n", "indian 0 -0.7128756 0.1011587 -0.8777275 -0.0004380795\n", "japanese 0 0.1288697 0.9425626 -0.2380748 0.3373437611\n", "korean 0 -0.2445193 -0.4744318 -0.8957870 -0.0048784496\n", "thai 0 1.5365848 0.1333256 0.2196970 -0.0113078024\n", " beef beef_broth beef_liver beer beet\n", "indian -0.7985278 0.2430186 -0.035598065 -0.002173738 0.01005813\n", "japanese 0.2241875 -0.3653020 -0.139551027 0.128905553 0.04923911\n", "korean 0.5366515 -0.6153237 0.213455197 -0.010828645 0.27325423\n", "thai 0.1570012 -0.9364154 -0.008032213 -0.035063746 -0.28279823\n", " bell_pepper bergamot berry bitter_orange black_bean\n", "indian 0.49074330 0 0.58947607 0.191256164 -0.1945233\n", "japanese 0.09074167 0 -0.25917977 -0.118915977 -0.3442400\n", "korean -0.57876763 0 -0.07874180 -0.007729435 -0.5220672\n", "thai 0.92554006 0 -0.07210196 -0.002983296 -0.4614426\n", " black_currant black_mustard_seed_oil black_pepper black_raspberry\n", "indian 0 0.38935801 -0.4453495 0\n", "japanese 0 -0.05452887 -0.5440869 0\n", "korean 0 -0.03929970 0.8025454 0\n", "thai 0 -0.21498372 -0.9854806 0\n", " black_sesame_seed black_tea blackberry blackberry_brandy\n", "indian -0.2759246 0.3079977 0.191256164 0\n", "japanese -0.6101687 -0.1671913 -0.118915977 0\n", "korean 1.5197674 -0.3036261 -0.007729435 0\n", "thai -0.1755656 -0.1487033 -0.002983296 0\n", " blue_cheese blueberry bone_oil bourbon_whiskey brandy\n", "indian 0 0.216164294 -0.2276744 0 0.22427587\n", "japanese 0 -0.119186087 0.3913019 0 -0.15595599\n", "korean 0 -0.007821986 0.2854487 0 -0.02562342\n", "thai 0 -0.004947048 -0.0253658 0 -0.05715244\n", "\n", "...\n", "and 308 more lines." ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 1000 }, "id": "GMbdfVmTKkJI", "outputId": "adf9ebdf-d69d-4a64-e9fd-e06e5322292e" } }, { "cell_type": "markdown", "source": [ "Utdataene viser koeffisientene som modellen lærte under treningen.\n", "\n", "### Evaluere den trente modellen\n", "\n", "Det er på tide å se hvordan modellen presterte 📏 ved å evaluere den på et testsett! La oss starte med å gjøre prediksjoner på testsettet.\n" ], "metadata": { "id": "tt2BfOxrKmcJ" } }, { "cell_type": "code", "execution_count": 9, "source": [ "# Make predictions on the test set\n", "results <- cuisines_test %>% select(cuisine) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test))\n", "\n", "# Print out results\n", "results %>% \n", " slice_head(n = 5)" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " cuisine .pred_class\n", "1 indian thai \n", "2 indian indian \n", "3 indian indian \n", "4 indian indian \n", "5 indian indian " ], "text/markdown": [ "\n", "A tibble: 5 × 2\n", "\n", "| cuisine <fct> | .pred_class <fct> |\n", "|---|---|\n", "| indian | thai |\n", "| indian | indian |\n", "| indian | indian |\n", "| indian | indian |\n", "| indian | indian |\n", "\n" ], "text/latex": [ "A tibble: 5 × 2\n", "\\begin{tabular}{ll}\n", " cuisine & .pred\\_class\\\\\n", " & \\\\\n", "\\hline\n", "\t indian & thai \\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\t indian & indian\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 2
cuisine.pred_class
<fct><fct>
indianthai
indianindian
indianindian
indianindian
indianindian
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 248 }, "id": "CqtckvtsKqax", "outputId": "e57fe557-6a68-4217-fe82-173328c5436d" } }, { "cell_type": "markdown", "source": [ "Flott jobb! I Tidymodels kan evaluering av modellprestasjon gjøres ved hjelp av [yardstick](https://yardstick.tidymodels.org/) - en pakke som brukes til å måle effektiviteten av modeller ved hjelp av ytelsesmetrikker. Som vi gjorde i vår leksjon om logistisk regresjon, la oss begynne med å beregne en forvirringsmatrise.\n" ], "metadata": { "id": "8w5N6XsBKss7" } }, { "cell_type": "code", "execution_count": 10, "source": [ "# Confusion matrix for categorical data\n", "conf_mat(data = results, truth = cuisine, estimate = .pred_class)\n" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " Truth\n", "Prediction chinese indian japanese korean thai\n", " chinese 83 1 8 15 10\n", " indian 4 163 1 2 6\n", " japanese 21 5 73 25 1\n", " korean 15 0 11 191 0\n", " thai 10 11 3 7 70" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 133 }, "id": "YvODvsLkK0iG", "outputId": "bb69da84-1266-47ad-b174-d43b88ca2988" } }, { "cell_type": "markdown", "source": [], "metadata": { "id": "c0HfPL16Lr6U" } }, { "cell_type": "code", "execution_count": 11, "source": [ "update_geom_defaults(geom = \"tile\", new = list(color = \"black\", alpha = 0.7))\n", "# Visualize confusion matrix\n", "results %>% \n", " conf_mat(cuisine, .pred_class) %>% \n", " autoplot(type = \"heatmap\")" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ "plot without title" ], "image/png": "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" }, "metadata": { "image/png": { "width": 420, "height": 420 } } } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 436 }, "id": "HsAtwukyLsvt", "outputId": "3032a224-a2c8-4270-b4f2-7bb620317400" } }, { "cell_type": "markdown", "source": [ "De mørkere rutene i forvirringsmatrisen indikerer et høyt antall tilfeller, og du kan forhåpentligvis se en diagonal linje med mørkere ruter som viser tilfeller der den predikerte og faktiske etiketten er den samme.\n", "\n", "La oss nå beregne oppsummeringsstatistikk for forvirringsmatrisen.\n" ], "metadata": { "id": "oOJC87dkLwPr" } }, { "cell_type": "code", "execution_count": 12, "source": [ "# Summary stats for confusion matrix\n", "conf_mat(data = results, truth = cuisine, estimate = .pred_class) %>% \n", "summary()" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " .metric .estimator .estimate\n", "1 accuracy multiclass 0.7880435\n", "2 kap multiclass 0.7276583\n", "3 sens macro 0.7780927\n", "4 spec macro 0.9477598\n", "5 ppv macro 0.7585583\n", "6 npv macro 0.9460080\n", "7 mcc multiclass 0.7292724\n", "8 j_index macro 0.7258524\n", "9 bal_accuracy macro 0.8629262\n", "10 detection_prevalence macro 0.2000000\n", "11 precision macro 0.7585583\n", "12 recall macro 0.7780927\n", "13 f_meas macro 0.7641862" ], "text/markdown": [ "\n", "A tibble: 13 × 3\n", "\n", "| .metric <chr> | .estimator <chr> | .estimate <dbl> |\n", "|---|---|---|\n", "| accuracy | multiclass | 0.7880435 |\n", "| kap | multiclass | 0.7276583 |\n", "| sens | macro | 0.7780927 |\n", "| spec | macro | 0.9477598 |\n", "| ppv | macro | 0.7585583 |\n", "| npv | macro | 0.9460080 |\n", "| mcc | multiclass | 0.7292724 |\n", "| j_index | macro | 0.7258524 |\n", "| bal_accuracy | macro | 0.8629262 |\n", "| detection_prevalence | macro | 0.2000000 |\n", "| precision | macro | 0.7585583 |\n", "| recall | macro | 0.7780927 |\n", "| f_meas | macro | 0.7641862 |\n", "\n" ], "text/latex": [ "A tibble: 13 × 3\n", "\\begin{tabular}{lll}\n", " .metric & .estimator & .estimate\\\\\n", " & & \\\\\n", "\\hline\n", "\t accuracy & multiclass & 0.7880435\\\\\n", "\t kap & multiclass & 0.7276583\\\\\n", "\t sens & macro & 0.7780927\\\\\n", "\t spec & macro & 0.9477598\\\\\n", "\t ppv & macro & 0.7585583\\\\\n", "\t npv & macro & 0.9460080\\\\\n", "\t mcc & multiclass & 0.7292724\\\\\n", "\t j\\_index & macro & 0.7258524\\\\\n", "\t bal\\_accuracy & macro & 0.8629262\\\\\n", "\t detection\\_prevalence & macro & 0.2000000\\\\\n", "\t precision & macro & 0.7585583\\\\\n", "\t recall & macro & 0.7780927\\\\\n", "\t f\\_meas & macro & 0.7641862\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 13 × 3
.metric.estimator.estimate
<chr><chr><dbl>
accuracy multiclass0.7880435
kap multiclass0.7276583
sens macro 0.7780927
spec macro 0.9477598
ppv macro 0.7585583
npv macro 0.9460080
mcc multiclass0.7292724
j_index macro 0.7258524
bal_accuracy macro 0.8629262
detection_prevalencemacro 0.2000000
precision macro 0.7585583
recall macro 0.7780927
f_meas macro 0.7641862
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 494 }, "id": "OYqetUyzL5Wz", "outputId": "6a84d65e-113d-4281-dfc1-16e8b70f37e6" } }, { "cell_type": "markdown", "source": [ "Hvis vi ser nærmere på noen målinger som nøyaktighet, sensitivitet, ppv, er vi ikke så verst i gang 🥳!\n", "\n", "## 4. Gå dypere\n", "\n", "La oss stille et subtilt spørsmål: Hvilke kriterier brukes for å avgjøre hvilken type mat som blir den predikerte utfallet?\n", "\n", "Vel, statistiske maskinlæringsalgoritmer, som logistisk regresjon, er basert på `sannsynlighet`; så det som faktisk blir predikert av en klassifiserer er en sannsynlighetsfordeling over et sett av mulige utfall. Klassen med høyest sannsynlighet blir deretter valgt som det mest sannsynlige utfallet for de gitte observasjonene.\n", "\n", "La oss se dette i praksis ved å lage både harde klasseprediksjoner og sannsynligheter.\n" ], "metadata": { "id": "43t7vz8vMJtW" } }, { "cell_type": "code", "execution_count": 13, "source": [ "# Make hard class prediction and probabilities\n", "results_prob <- cuisines_test %>%\n", " select(cuisine) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test)) %>% \n", " bind_cols(mr_fit %>% predict(new_data = cuisines_test, type = \"prob\"))\n", "\n", "# Print out results\n", "results_prob %>% \n", " slice_head(n = 5)" ], "outputs": [ { "output_type": "display_data", "data": { "text/plain": [ " cuisine .pred_class .pred_chinese .pred_indian .pred_japanese .pred_korean\n", "1 indian thai 1.551259e-03 0.4587877 5.988039e-04 2.428503e-04\n", "2 indian indian 2.637133e-05 0.9999488 6.648651e-07 2.259993e-05\n", "3 indian indian 1.049433e-03 0.9909982 1.060937e-03 1.644947e-05\n", "4 indian indian 6.237482e-02 0.4763035 9.136702e-02 3.660913e-01\n", "5 indian indian 1.431745e-02 0.9418551 2.945239e-02 8.721782e-03\n", " .pred_thai \n", "1 5.388194e-01\n", "2 1.577948e-06\n", "3 6.874989e-03\n", "4 3.863391e-03\n", "5 5.653283e-03" ], "text/markdown": [ "\n", "A tibble: 5 × 7\n", "\n", "| cuisine <fct> | .pred_class <fct> | .pred_chinese <dbl> | .pred_indian <dbl> | .pred_japanese <dbl> | .pred_korean <dbl> | .pred_thai <dbl> |\n", "|---|---|---|---|---|---|---|\n", "| indian | thai | 1.551259e-03 | 0.4587877 | 5.988039e-04 | 2.428503e-04 | 5.388194e-01 |\n", "| indian | indian | 2.637133e-05 | 0.9999488 | 6.648651e-07 | 2.259993e-05 | 1.577948e-06 |\n", "| indian | indian | 1.049433e-03 | 0.9909982 | 1.060937e-03 | 1.644947e-05 | 6.874989e-03 |\n", "| indian | indian | 6.237482e-02 | 0.4763035 | 9.136702e-02 | 3.660913e-01 | 3.863391e-03 |\n", "| indian | indian | 1.431745e-02 | 0.9418551 | 2.945239e-02 | 8.721782e-03 | 5.653283e-03 |\n", "\n" ], "text/latex": [ "A tibble: 5 × 7\n", "\\begin{tabular}{lllllll}\n", " cuisine & .pred\\_class & .pred\\_chinese & .pred\\_indian & .pred\\_japanese & .pred\\_korean & .pred\\_thai\\\\\n", " & & & & & & \\\\\n", "\\hline\n", "\t indian & thai & 1.551259e-03 & 0.4587877 & 5.988039e-04 & 2.428503e-04 & 5.388194e-01\\\\\n", "\t indian & indian & 2.637133e-05 & 0.9999488 & 6.648651e-07 & 2.259993e-05 & 1.577948e-06\\\\\n", "\t indian & indian & 1.049433e-03 & 0.9909982 & 1.060937e-03 & 1.644947e-05 & 6.874989e-03\\\\\n", "\t indian & indian & 6.237482e-02 & 0.4763035 & 9.136702e-02 & 3.660913e-01 & 3.863391e-03\\\\\n", "\t indian & indian & 1.431745e-02 & 0.9418551 & 2.945239e-02 & 8.721782e-03 & 5.653283e-03\\\\\n", "\\end{tabular}\n" ], "text/html": [ "\n", "\n", "\n", "\t\n", "\t\n", "\n", "\n", "\t\n", "\t\n", "\t\n", "\t\n", "\t\n", "\n", "
A tibble: 5 × 7
cuisine.pred_class.pred_chinese.pred_indian.pred_japanese.pred_korean.pred_thai
<fct><fct><dbl><dbl><dbl><dbl><dbl>
indianthai 1.551259e-030.45878775.988039e-042.428503e-045.388194e-01
indianindian2.637133e-050.99994886.648651e-072.259993e-051.577948e-06
indianindian1.049433e-030.99099821.060937e-031.644947e-056.874989e-03
indianindian6.237482e-020.47630359.136702e-023.660913e-013.863391e-03
indianindian1.431745e-020.94185512.945239e-028.721782e-035.653283e-03
\n" ] }, "metadata": {} } ], "metadata": { "colab": { "base_uri": "https://localhost:8080/", "height": 248 }, "id": "xdKNs-ZPMTJL", "outputId": "68f6ac5a-725a-4eff-9ea6-481fef00e008" } }, { "cell_type": "markdown", "source": [ "✅ Kan du forklare hvorfor modellen er ganske sikker på at den første observasjonen er thai?\n", "\n", "## **🚀Utfordring**\n", "\n", "I denne leksjonen brukte du de rensede dataene dine til å bygge en maskinlæringsmodell som kan forutsi en nasjonal matkultur basert på en rekke ingredienser. Ta deg tid til å lese gjennom de [mange alternativene](https://www.tidymodels.org/find/parsnip/#models) som Tidymodels tilbyr for å klassifisere data og [andre måter](https://parsnip.tidymodels.org/articles/articles/Examples.html#multinom_reg-models) å tilpasse multinomial regresjon.\n", "\n", "#### TAKK TIL:\n", "\n", "[`Allison Horst`](https://twitter.com/allison_horst/) for å lage de fantastiske illustrasjonene som gjør R mer innbydende og engasjerende. Finn flere illustrasjoner i hennes [galleri](https://www.google.com/url?q=https://github.com/allisonhorst/stats-illustrations&sa=D&source=editors&ust=1626380772530000&usg=AOvVaw3zcfyCizFQZpkSLzxiiQEM).\n", "\n", "[Cassie Breviu](https://www.twitter.com/cassieview) og [Jen Looper](https://www.twitter.com/jenlooper) for å lage den originale Python-versjonen av denne modulen ♥️\n", "\n", "
\n", "Skulle gjerne ha slengt inn noen vitser, men jeg skjønner ikke matordspill 😅.\n", "\n", "
\n", "\n", "Lykke til med læringen,\n", "\n", "[Eric](https://twitter.com/ericntay), Gold Microsoft Learn Student Ambassador.\n" ], "metadata": { "id": "2tWVHMeLMYdM" } }, { "cell_type": "markdown", "metadata": {}, "source": [ "\n---\n\n**Ansvarsfraskrivelse**: \nDette dokumentet er oversatt ved hjelp av AI-oversettelsestjenesten [Co-op Translator](https://github.com/Azure/co-op-translator). Selv om vi tilstreber nøyaktighet, vennligst vær oppmerksom på at automatiske oversettelser kan inneholde feil eller unøyaktigheter. Det originale dokumentet på sitt opprinnelige språk bør anses som den autoritative kilden. For kritisk informasjon anbefales profesjonell menneskelig oversettelse. Vi er ikke ansvarlige for eventuelle misforståelser eller feiltolkninger som oppstår ved bruk av denne oversettelsen.\n" ] } ] }