{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "\n# Imputing missing values before building an estimator\n\nMissing values can be replaced by the mean, the median or the most frequent\nvalue using the basic :class:`~sklearn.impute.SimpleImputer`.\n\nIn this example we will investigate different imputation techniques:\n\n- imputation by the constant value 0\n- imputation by the mean value of each feature\n- k nearest neighbor imputation\n- iterative imputation\n\nIn all the cases, for each feature, we add a new feature indicating the missingness.\n\nWe will use two datasets: Diabetes dataset which consists of 10 feature\nvariables collected from diabetes patients with an aim to predict disease\nprogression and California housing dataset for which the target is the median\nhouse value for California districts.\n\nAs neither of these datasets have missing values, we will remove some\nvalues to create new versions with artificially missing data. The performance\nof\n:class:`~sklearn.ensemble.RandomForestRegressor` on the full original dataset\nis then compared the performance on the altered datasets with the artificially\nmissing values imputed using different techniques.\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "# Authors: The scikit-learn developers\n# SPDX-License-Identifier: BSD-3-Clause" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Download the data and make missing values sets\n\nFirst we download the two datasets. Diabetes dataset is shipped with\nscikit-learn. It has 442 entries, each with 10 features. California housing\ndataset is much larger with 20640 entries and 8 features. It needs to be\ndownloaded. We will only use the first 300 entries for the sake of speeding\nup the calculations but feel free to use the whole dataset.\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import numpy as np\n\nfrom sklearn.datasets import fetch_california_housing, load_diabetes\n\nX_diabetes, y_diabetes = load_diabetes(return_X_y=True)\nX_california, y_california = fetch_california_housing(return_X_y=True)\n\nX_diabetes = X_diabetes[:300]\ny_diabetes = y_diabetes[:300]\nX_california = X_california[:300]\ny_california = y_california[:300]\n\n\ndef add_missing_values(X_full, y_full, rng):\n n_samples, n_features = X_full.shape\n\n # Add missing values in 75% of the lines\n missing_rate = 0.75\n n_missing_samples = int(n_samples * missing_rate)\n\n missing_samples = np.zeros(n_samples, dtype=bool)\n missing_samples[:n_missing_samples] = True\n\n rng.shuffle(missing_samples)\n missing_features = rng.randint(0, n_features, n_missing_samples)\n X_missing = X_full.copy()\n X_missing[missing_samples, missing_features] = np.nan\n y_missing = y_full.copy()\n\n return X_missing, y_missing\n\n\nrng = np.random.RandomState(42)\nX_miss_diabetes, y_miss_diabetes = add_missing_values(X_diabetes, y_diabetes, rng)\nX_miss_california, y_miss_california = add_missing_values(\n X_california, y_california, rng\n)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Impute the missing data and score\nNow we will write a function which will score the results on the differently\nimputed data, including the case of no imputation for full data.\nWe will use :class:`~sklearn.ensemble.RandomForestRegressor` for the target\nregression.\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "from sklearn.ensemble import RandomForestRegressor\n\n# To use the experimental IterativeImputer, we need to explicitly ask for it:\nfrom sklearn.experimental import enable_iterative_imputer # noqa: F401\nfrom sklearn.impute import IterativeImputer, KNNImputer, SimpleImputer\nfrom sklearn.model_selection import cross_val_score\nfrom sklearn.pipeline import make_pipeline\nfrom sklearn.preprocessing import RobustScaler\n\nN_SPLITS = 4\n\n\ndef get_score(X, y, imputer=None):\n regressor = RandomForestRegressor(random_state=0)\n if imputer is not None:\n estimator = make_pipeline(imputer, regressor)\n else:\n estimator = regressor\n scores = cross_val_score(\n estimator, X, y, scoring=\"neg_mean_squared_error\", cv=N_SPLITS\n )\n return scores.mean(), scores.std()\n\n\nx_labels = []\n\nmses_diabetes = np.zeros(5)\nstds_diabetes = np.zeros(5)\nmses_california = np.zeros(5)\nstds_california = np.zeros(5)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Estimate the score\nFirst, we want to estimate the score on the original data:\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "mses_diabetes[0], stds_diabetes[0] = get_score(X_diabetes, y_diabetes)\nmses_california[0], stds_california[0] = get_score(X_california, y_california)\nx_labels.append(\"Full Data\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Replace missing values by 0\n\nNow we will estimate the score on the data where the missing values are\nreplaced by 0:\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "imputer = SimpleImputer(strategy=\"constant\", fill_value=0, add_indicator=True)\nmses_diabetes[1], stds_diabetes[1] = get_score(\n X_miss_diabetes, y_miss_diabetes, imputer\n)\nmses_california[1], stds_california[1] = get_score(\n X_miss_california, y_miss_california, imputer\n)\nx_labels.append(\"Zero Imputation\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Impute missing values with mean\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "imputer = SimpleImputer(strategy=\"mean\", add_indicator=True)\nmses_diabetes[2], stds_diabetes[2] = get_score(\n X_miss_diabetes, y_miss_diabetes, imputer\n)\nmses_california[2], stds_california[2] = get_score(\n X_miss_california, y_miss_california, imputer\n)\nx_labels.append(\"Mean Imputation\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### kNN-imputation of the missing values\n\n:class:`~sklearn.impute.KNNImputer` imputes missing values using the weighted\nor unweighted mean of the desired number of nearest neighbors. If your features\nhave vastly different scales (as in the California housing dataset),\nconsider re-scaling them to potentially improve performance.\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "imputer = KNNImputer(add_indicator=True)\nmses_diabetes[3], stds_diabetes[3] = get_score(\n X_miss_diabetes, y_miss_diabetes, imputer\n)\nmses_california[3], stds_california[3] = get_score(\n X_miss_california, y_miss_california, make_pipeline(RobustScaler(), imputer)\n)\nx_labels.append(\"KNN Imputation\")" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "### Iterative imputation of the missing values\n\nAnother option is the :class:`~sklearn.impute.IterativeImputer`. This uses\nround-robin regression, modeling each feature with missing values as a\nfunction of other features, in turn. We use the class's default choice\nof the regressor model (:class:`~sklearn.linear_model.BayesianRidge`)\nto predict missing feature values. The performance of the predictor\nmay be negatively affected by vastly different scales of the features,\nso we re-scale the features in the California housing dataset.\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "imputer = IterativeImputer(add_indicator=True)\n\nmses_diabetes[4], stds_diabetes[4] = get_score(\n X_miss_diabetes, y_miss_diabetes, imputer\n)\nmses_california[4], stds_california[4] = get_score(\n X_miss_california, y_miss_california, make_pipeline(RobustScaler(), imputer)\n)\nx_labels.append(\"Iterative Imputation\")\n\nmses_diabetes = mses_diabetes * -1\nmses_california = mses_california * -1" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "## Plot the results\n\nFinally we are going to visualize the score:\n\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import matplotlib.pyplot as plt\n\nn_bars = len(mses_diabetes)\nxval = np.arange(n_bars)\n\ncolors = [\"r\", \"g\", \"b\", \"orange\", \"black\"]\n\n# plot diabetes results\nplt.figure(figsize=(12, 6))\nax1 = plt.subplot(121)\nfor j in xval:\n ax1.barh(\n j,\n mses_diabetes[j],\n xerr=stds_diabetes[j],\n color=colors[j],\n alpha=0.6,\n align=\"center\",\n )\n\nax1.set_title(\"Imputation Techniques with Diabetes Data\")\nax1.set_xlim(left=np.min(mses_diabetes) * 0.9, right=np.max(mses_diabetes) * 1.1)\nax1.set_yticks(xval)\nax1.set_xlabel(\"MSE\")\nax1.invert_yaxis()\nax1.set_yticklabels(x_labels)\n\n# plot california dataset results\nax2 = plt.subplot(122)\nfor j in xval:\n ax2.barh(\n j,\n mses_california[j],\n xerr=stds_california[j],\n color=colors[j],\n alpha=0.6,\n align=\"center\",\n )\n\nax2.set_title(\"Imputation Techniques with California Data\")\nax2.set_yticks(xval)\nax2.set_xlabel(\"MSE\")\nax2.invert_yaxis()\nax2.set_yticklabels([\"\"] * n_bars)\n\nplt.show()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "You can also try different techniques. For instance, the median is a more\nrobust estimator for data with high magnitude variables which could dominate\nresults (otherwise known as a 'long tail').\n\n" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.11.14" } }, "nbformat": 4, "nbformat_minor": 0 }