{ "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": "2025-09-06T14:34:22+00:00", "source_file": "4-Classification/2-Classifiers-1/solution/R/lesson_11-R.ipynb", "language_code": "sk" } }, "cells": [ { "cell_type": "markdown", "source": [], "metadata": { "id": "zs2woWv_HoE8" } }, { "cell_type": "markdown", "source": [ "## Klasifikátory kuchýň 1\n", "\n", "V tejto lekcii preskúmame rôzne klasifikátory na *predpovedanie národnej kuchyne na základe skupiny ingrediencií.* Pri tom sa dozvieme viac o spôsoboch, akými môžu byť algoritmy využívané na úlohy klasifikácie.\n", "\n", "### [**Kvíz pred prednáškou**](https://gray-sand-07a10f403.1.azurestaticapps.net/quiz/21/)\n", "\n", "### **Príprava**\n", "\n", "Táto lekcia nadväzuje na našu [predchádzajúcu lekciu](https://github.com/microsoft/ML-For-Beginners/blob/main/4-Classification/1-Introduction/solution/lesson_10-R.ipynb), kde sme:\n", "\n", "- Jemne uviedli klasifikácie pomocou datasetu o všetkých úžasných kuchyniach Ázie a Indie 😋.\n", "\n", "- Preskúmali niektoré [slovesá dplyr](https://dplyr.tidyverse.org/) na prípravu a čistenie našich dát.\n", "\n", "- Vytvorili krásne vizualizácie pomocou ggplot2.\n", "\n", "- Ukázali, ako sa vysporiadať s nevyváženými dátami ich predspracovaním pomocou [recipes](https://recipes.tidymodels.org/articles/Simple_Example.html).\n", "\n", "- Demonštrovali, ako `prep` a `bake` náš recept, aby sme si overili, že funguje podľa očakávaní.\n", "\n", "#### **Predpoklady**\n", "\n", "Na túto lekciu budeme potrebovať nasledujúce balíky na čistenie, prípravu a vizualizáciu našich dát:\n", "\n", "- `tidyverse`: [tidyverse](https://www.tidyverse.org/) je [kolekcia balíkov pre R](https://www.tidyverse.org/packages), ktorá robí dátovú vedu rýchlejšou, jednoduchšou a zábavnejšou!\n", "\n", "- `tidymodels`: [tidymodels](https://www.tidymodels.org/) je rámec [kolekcie balíkov](https://www.tidymodels.org/packages/) na modelovanie a strojové učenie.\n", "\n", "- `themis`: [balík themis](https://themis.tidymodels.org/) poskytuje dodatočné kroky receptov na riešenie nevyvážených dát.\n", "\n", "- `nnet`: [balík nnet](https://cran.r-project.org/web/packages/nnet/nnet.pdf) poskytuje funkcie na odhadovanie dopredných neurónových sietí s jednou skrytou vrstvou a na modely multinomiálnej logistickej regresie.\n", "\n", "Môžete ich nainštalovať takto:\n" ], "metadata": { "id": "iDFOb3ebHwQC" } }, { "cell_type": "markdown", "source": [ "`install.packages(c(\"tidyverse\", \"tidymodels\", \"DataExplorer\", \"here\"))`\n", "\n", "Alternatívne, nasledujúci skript skontroluje, či máte nainštalované balíky potrebné na dokončenie tohto modulu, a v prípade, že chýbajú, ich nainštaluje za vás.\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. Rozdeľte údaje na tréningovú a testovaciu množinu.\n", "\n", "Začneme výberom niekoľkých krokov z našej predchádzajúcej lekcie.\n", "\n", "### Odstráňte najbežnejšie ingrediencie, ktoré spôsobujú zmätok medzi rôznymi kuchyňami, pomocou `dplyr::select()`.\n", "\n", "Každý miluje ryžu, cesnak a zázvor!\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": [ "Perfektné! Teraz je čas rozdeliť údaje tak, aby 70 % údajov išlo na tréning a 30 % na testovanie. Pri rozdeľovaní údajov použijeme aj techniku `stratifikácie`, aby sme `zachovali pomer jednotlivých kuchýň` v tréningových a validačných datasetoch.\n", "\n", "[rsample](https://rsample.tidymodels.org/), balík v Tidymodels, poskytuje infraštruktúru na efektívne rozdeľovanie a resampling údajov:\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. Riešenie nevyvážených údajov\n", "\n", "Ako ste si mohli všimnúť v pôvodnej dátovej sade, ako aj v našej tréningovej sade, existuje pomerne nerovnomerné rozdelenie počtu kuchýň. Kórejské kuchyne sú *takmer* 3-krát častejšie ako thajské kuchyne. Nevývážené údaje často negatívne ovplyvňujú výkon modelu. Mnohé modely dosahujú najlepšie výsledky, keď je počet pozorovaní rovnaký, a preto majú tendenciu mať problémy s nevyváženými údajmi.\n", "\n", "Existujú dva hlavné spôsoby, ako riešiť nevyvážené dátové sady:\n", "\n", "- pridanie pozorovaní do menšinovej triedy: `Over-sampling`, napríklad pomocou algoritmu SMOTE, ktorý synteticky generuje nové príklady menšinovej triedy pomocou najbližších susedov týchto prípadov.\n", "\n", "- odstránenie pozorovaní z väčšinovej triedy: `Under-sampling`\n", "\n", "V našej predchádzajúcej lekcii sme ukázali, ako riešiť nevyvážené dátové sady pomocou `receptu`. Recept si môžeme predstaviť ako plán, ktorý popisuje, aké kroky by sa mali aplikovať na dátovú sadu, aby bola pripravená na analýzu údajov. V našom prípade chceme dosiahnuť rovnomerné rozdelenie počtu našich kuchýň pre našu `tréningovú sadu`. Poďme sa do toho pustiť.\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": [ "Môžete samozrejme potvrdiť (pomocou prípravy a pečenia), že recept bude fungovať tak, ako očakávate - všetky označenia kuchyne majú `559` pozorovaní.\n", "\n", "Keďže tento recept budeme používať ako predspracovanie pre modelovanie, `workflow()` za nás vykoná všetku prípravu a pečenie, takže recept nebudeme musieť manuálne odhadovať.\n", "\n", "Teraz sme pripravení trénovať model 👩‍💻👨‍💻!\n", "\n", "## 3. Výber klasifikátora\n", "\n", "

\n", " \n", "

Ilustrácia od @allison_horst
\n" ], "metadata": { "id": "NBL3PqIWJBBB" } }, { "cell_type": "markdown", "source": [ "Teraz musíme rozhodnúť, ktorý algoritmus použiť na túto úlohu 🤔.\n", "\n", "V Tidymodels poskytuje [`parsnip package`](https://parsnip.tidymodels.org/index.html) konzistentné rozhranie na prácu s modelmi naprieč rôznymi enginmi (balíčkami). Pozrite si dokumentáciu k parsnip, aby ste preskúmali [typy modelov a enginy](https://www.tidymodels.org/find/parsnip/#models) a ich zodpovedajúce [argumenty modelov](https://www.tidymodels.org/find/parsnip/#model-args). Rozmanitosť môže byť na prvý pohľad dosť mätúca. Napríklad nasledujúce metódy zahŕňajú techniky klasifikácie:\n", "\n", "- C5.0 modely založené na pravidlách\n", "- Flexibilné diskriminačné modely\n", "- Lineárne diskriminačné modely\n", "- Regularizované diskriminačné modely\n", "- Modely logistickej regresie\n", "- Modely multinomiálnej regresie\n", "- Modely naivného Bayesa\n", "- Podporné vektorové stroje\n", "- Najbližší susedia\n", "- Rozhodovacie stromy\n", "- Ensemble metódy\n", "- Neurónové siete\n", "\n", "A zoznam pokračuje!\n", "\n", "### **Aký klasifikátor zvoliť?**\n", "\n", "Takže, ktorý klasifikátor by ste si mali vybrať? Často je dobrým spôsobom testovania prejsť viacerými a hľadať dobrý výsledok.\n", "\n", "> AutoML tento problém elegantne rieši tým, že vykonáva tieto porovnania v cloude, čo vám umožňuje vybrať najlepší algoritmus pre vaše dáta. Vyskúšajte to [tu](https://docs.microsoft.com/learn/modules/automate-model-selection-with-azure-automl/?WT.mc_id=academic-77952-leestott)\n", "\n", "Výber klasifikátora však závisí aj od nášho problému. Napríklad, keď výsledok môže byť kategorizovaný do `viac ako dvoch tried`, ako v našom prípade, musíte použiť `algoritmus pre multiklasifikáciu` namiesto `binárnej klasifikácie.`\n", "\n", "### **Lepší prístup**\n", "\n", "Lepším spôsobom ako náhodne hádať je však riadiť sa nápadmi z tohto stiahnuteľného [ML Cheat sheet](https://docs.microsoft.com/azure/machine-learning/algorithm-cheat-sheet?WT.mc_id=academic-77952-leestott). Tu zistíme, že pre náš problém s multiklasifikáciou máme niekoľko možností:\n", "\n", "

\n", " \n", "

Časť Microsoftovho prehľadu algoritmov, ktorá podrobne opisuje možnosti multiklasifikácie
\n" ], "metadata": { "id": "a6DLAZ3vJZ14" } }, { "cell_type": "markdown", "source": [ "### **Úvaha**\n", "\n", "Pozrime sa, či dokážeme logicky zhodnotiť rôzne prístupy vzhľadom na obmedzenia, ktoré máme:\n", "\n", "- **Hlboké neurónové siete sú príliš náročné**. Vzhľadom na náš čistý, ale minimálny dataset a fakt, že trénovanie prebieha lokálne cez notebooky, sú hlboké neurónové siete pre túto úlohu príliš náročné.\n", "\n", "- **Žiadny dvojtriedny klasifikátor**. Nepoužívame dvojtriedny klasifikátor, takže možnosť one-vs-all je vylúčená.\n", "\n", "- **Rozhodovací strom alebo logistická regresia by mohli fungovať**. Rozhodovací strom by mohol fungovať, rovnako ako multinomiálna regresia/multitriedna logistická regresia pre multitriedne dáta.\n", "\n", "- **Multitriedne Boosted Decision Trees riešia iný problém**. Multitriedny Boosted Decision Tree je najvhodnejší pre neparametrické úlohy, napríklad úlohy zamerané na vytváranie rebríčkov, takže pre nás nie je užitočný.\n", "\n", "Okrem toho, predtým než sa pustíme do zložitejších modelov strojového učenia, ako sú ensemble metódy, je zvyčajne dobré začať s najjednoduchším možným modelom, aby sme získali predstavu o tom, čo sa deje. Preto v tejto lekcii začneme s modelom `multinomiálnej regresie`.\n", "\n", "> Logistická regresia je technika používaná, keď je výstupná premenná kategóriálna (alebo nominálna). Pri binárnej logistickej regresii je počet výstupných premenných dva, zatiaľ čo pri multinomiálnej logistickej regresii je počet výstupných premenných viac ako dva. Viac informácií nájdete v [Pokročilé regresné metódy](https://bookdown.org/chua/ber642_advanced_regression/multinomial-logistic-regression.html).\n", "\n", "## 4. Trénovanie a hodnotenie modelu multinomiálnej logistickej regresie\n", "\n", "V Tidymodels, `parsnip::multinom_reg()`, definuje model, ktorý používa lineárne prediktory na predpovedanie multitriednych dát pomocou multinomiálneho rozdelenia. Pozrite si `?multinom_reg()` pre rôzne spôsoby/enginy, ktoré môžete použiť na fitovanie tohto modelu.\n", "\n", "V tomto príklade budeme fitovať model multinomiálnej regresie cez predvolený engine [nnet](https://cran.r-project.org/web/packages/nnet/nnet.pdf).\n", "\n", "> Hodnotu pre `penalty` som vybral tak trochu náhodne. Existujú lepšie spôsoby, ako túto hodnotu zvoliť, napríklad pomocou `resamplingu` a `ladenia` modelu, o ktorých budeme hovoriť neskôr.\n", ">\n", "> Pozrite si [Tidymodels: Začnite](https://www.tidymodels.org/start/tuning/), ak sa chcete dozvedieť viac o ladení hyperparametrov modelu.\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": [ "Skvelá práca 🥳! Teraz, keď máme recept a špecifikáciu modelu, musíme nájsť spôsob, ako ich spojiť do objektu, ktorý najprv predspracuje dáta, potom na predspracovaných dátach natrénuje model a zároveň umožní aj prípadné aktivity po spracovaní. V Tidymodels sa tento praktický objekt nazýva [`workflow`](https://workflows.tidymodels.org/) a pohodlne uchováva vaše modelovacie komponenty! Toto by sme v *Pythone* nazvali *pipelines*.\n", "\n", "Takže poďme všetko zabaliť do workflowu!📦\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": [ "Pracovné postupy 👌👌! **`workflow()`** môže byť nastavený podobne ako model. Takže, je čas trénovať model!\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": [ "Výstup zobrazuje koeficienty, ktoré model naučil počas tréningu.\n", "\n", "### Vyhodnotenie vytrénovaného modelu\n", "\n", "Je čas zistiť, ako si model viedol 📏, vyhodnotením na testovacej množine! Začnime tým, že urobíme predpovede na testovacej množine.\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": [ "Skvelá práca! V Tidymodels je hodnotenie výkonu modelu možné pomocou [yardstick](https://yardstick.tidymodels.org/) - balíka používaného na meranie efektívnosti modelov pomocou metrík výkonu. Ako sme to urobili v našej lekcii o logistickej regresii, začnime výpočtom matice zámien.\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": [ "Pri práci s viacerými triedami je vo všeobecnosti intuitívnejšie vizualizovať to ako tepelnú mapu, napríklad takto:\n" ], "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": [ "Tmavšie štvorce v grafe matice zámien naznačujú vysoký počet prípadov, a dúfajme, že vidíte diagonálnu líniu tmavších štvorcov, ktorá označuje prípady, kde predpovedaná a skutočná značka sú rovnaké.\n", "\n", "Teraz vypočítajme súhrnné štatistiky pre maticu zámien.\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": [ "Ak sa zameriame na niektoré metriky, ako presnosť, citlivosť, ppv, na začiatok na tom nie sme zle 🥳!\n", "\n", "## 4. Hlbšie skúmanie\n", "\n", "Položme si jednu jemnú otázku: Aké kritériá sa používajú na určenie konkrétneho typu kuchyne ako predpokladaného výsledku?\n", "\n", "No, štatistické algoritmy strojového učenia, ako logistická regresia, sú založené na `pravdepodobnosti`; takže to, čo klasifikátor skutočne predpovedá, je pravdepodobnostné rozdelenie nad množinou možných výsledkov. Trieda s najvyššou pravdepodobnosťou je potom vybraná ako najpravdepodobnejší výsledok pre dané pozorovania.\n", "\n", "Pozrime sa na to v praxi tým, že urobíme tvrdé predikcie tried a aj pravdepodobnosti.\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": [ "✅ Môžete vysvetliť, prečo si model celkom iste myslí, že prvé pozorovanie je thajské?\n", "\n", "## **🚀Výzva**\n", "\n", "V tejto lekcii ste použili svoje vyčistené dáta na vytvorenie modelu strojového učenia, ktorý dokáže predpovedať národnú kuchyňu na základe série ingrediencií. Nájdite si čas na preštudovanie [mnohých možností](https://www.tidymodels.org/find/parsnip/#models), ktoré Tidymodels ponúka na klasifikáciu dát, a [iných spôsobov](https://parsnip.tidymodels.org/articles/articles/Examples.html#multinom_reg-models), ako prispôsobiť multinomiálnu regresiu.\n", "\n", "#### POĎAKOVANIE:\n", "\n", "[`Allison Horst`](https://twitter.com/allison_horst/) za vytvorenie úžasných ilustrácií, ktoré robia R prístupnejším a pútavejším. Viac ilustrácií nájdete v jej [galérii](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) a [Jen Looper](https://www.twitter.com/jenlooper) za vytvorenie pôvodnej verzie tohto modulu v Pythone ♥️\n", "\n", "
\n", "Pridal by som nejaké vtipy, ale nerozumiem jedlým slovným hračkám 😅.\n", "\n", "
\n", "\n", "Šťastné učenie,\n", "\n", "[Eric](https://twitter.com/ericntay), Zlatý ambasádor Microsoft Learn.\n" ], "metadata": { "id": "2tWVHMeLMYdM" } }, { "cell_type": "markdown", "metadata": {}, "source": [ "\n---\n\n**Upozornenie**: \nTento dokument bol preložený pomocou služby na automatický preklad [Co-op Translator](https://github.com/Azure/co-op-translator). Aj keď sa snažíme o presnosť, upozorňujeme, že automatické preklady môžu obsahovať chyby alebo nepresnosti. Pôvodný dokument v jeho pôvodnom jazyku by mal byť považovaný za autoritatívny zdroj. Pre dôležité informácie sa odporúča profesionálny ľudský preklad. Nezodpovedáme za akékoľvek nedorozumenia alebo nesprávne interpretácie vyplývajúce z použitia tohto prekladu.\n" ] } ] }