{ "nbformat": 4, "nbformat_minor": 0, "metadata": { "colab": { "name": "Analysing Strava activities using Colab, Pandas and Matplotlib (Part 4)", "version": "0.3.2", "views": {}, "default_view": {}, "provenance": [ { "file_id": "1Png2gQvNaKJrq_YVhoFbkIoY9vzqq5sC", "timestamp": 1528233508683 }, { "file_id": "1eR2xoY-eOFodCIQuRROGytZrkioHLddt", "timestamp": 1527888313301 } ], "private_outputs": true, "collapsed_sections": [] }, "kernelspec": { "name": "python3", "display_name": "Python 3" }, "accelerator": "GPU" }, "cells": [ { "metadata": { "id": "PV6ioN1CXslD", "colab_type": "text" }, "cell_type": "markdown", "source": [ "# Preparation" ] }, { "metadata": { "id": "EHDKAkTztUbi", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Install gspread\n", "!pip install --upgrade -q gspread" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "E51hj0o2tUcN", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Import Strava activitities from Google Sheets\n", "\n", "from google.colab import auth\n", "auth.authenticate_user()\n", "\n", "import gspread\n", "from oauth2client.client import GoogleCredentials\n", "\n", "gc = gspread.authorize(GoogleCredentials.get_application_default())\n", "\n", "worksheet = gc.open('Strava 2018-07-02').sheet1\n", "\n", "# get_all_values gives a list of rows.\n", "rows = worksheet.get_all_values()\n", "\n", "# Convert to a DataFrame and render.\n", "import pandas as pd\n", "activities_raw = pd.DataFrame.from_records(rows[1:], columns=rows[0]);" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "CXFqnmhmt9wM", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Index activities, convert data types, convert units\n", "activities = activities_raw.copy()\n", "\n", "# Project\n", "activities = activities[['date', 'type', 'distance', 'elapsed_time', 'commute', 'name']]\n", "\n", "# Set data types\n", "activities['distance'] = activities['distance'].astype(float)\n", "activities['elapsed_time'] = activities['elapsed_time'].astype(float)\n", "activities['commute'] = activities['commute'].apply(lambda x: 1 if x == 'TRUE' else 0).astype(int)\n", "\n", "# Convert units\n", "activities['distance'] = activities['distance'] / 1000\n", "activities['elapsed_time'] = activities['elapsed_time'] / 60\n", "\n", "# Number of activities (at this aggregation level, trivially 1)\n", "activities['count'] = 1\n", "\n", "# Index\n", "activities['date'] = pd.to_datetime(activities_raw['date'])\n", "activities.set_index(['date', 'type'], inplace=True)" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "JZ9PgCHMHqK0", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Define formatters for readability\n", "formatters = {\n", " 'distance': '{:.0f}km'.format,\n", " 'elapsed_time':' {:.0f}min'.format,\n", "}\n", "\n", "detailed_formatters = {\n", " 'distance': '{:.1f}km'.format,\n", " 'elapsed_time':' {:.1f}min'.format,\n", "}" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "MFCgnpUiXraH", "colab_type": "text" }, "cell_type": "markdown", "source": [ "# Exploration" ] }, { "metadata": { "id": "KmrQVbsdawmR", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Print a few random activities\n", "print(\n", " activities[['distance', 'elapsed_time', 'commute', 'count']]\n", " .sample(5)\n", " .sort_index()\n", " .to_string(formatters=formatters))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "_71DHKJuH8z-", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Print totals\n", "print(\n", " activities\n", " .groupby(level='type')\n", " .sum()\n", " .sort_values(by='count', ascending=False)\n", " .to_string(formatters=formatters))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "2jiWJhR0hJv8", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Show date range\n", "\n", "first_date = activities.index.levels[0].min()\n", "last_date = activities.index.levels[0].max()\n", "print('Activities recorded from %s to %s, total of %d days.' % (\n", " first_date.strftime('%Y-%m-%d'),\n", " last_date.strftime('%Y-%m-%d'),\n", " (last_date - first_date).days))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "3t-tHRDxwa68", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Activities with the longest distance\n", "print(\n", " activities['distance']\n", " .groupby('type').nlargest(3)\n", " .to_string(float_format='%.1fkm'))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "n6rUBZJ40v0-", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Activities with the longest duration\n", "print(\n", " (activities['elapsed_time']\n", " .groupby('type').nlargest(4) / 60)\n", " .to_string(float_format='%.1fh'))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "ijelhc4oPasQ", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Summarize by day and activity type\n", "import numpy as np\n", "\n", "activities_by_day_type = (\n", " activities\n", " .groupby([pd.Grouper(freq='D', level=0), 'type'])\n", " .sum())\n", "\n", "activities_by_day_type = (\n", " activities_by_day_type\n", " .reset_index(level='type')\n", " .to_period('D')\n", " .reset_index()\n", " .set_index(['date', 'type']))\n", "\n", "first_day = activities_by_day_type.index.levels[0].min()\n", "last_day = activities_by_day_type.index.levels[0].max()\n", "days = pd.period_range(first_day, last_day, freq='D')\n", "types = activities_by_day_type.index.levels[1].unique()\n", "index = pd.MultiIndex.from_product([days, types], names=['day', 'type'])\n", "activities_by_day_type = activities_by_day_type.reindex(index, fill_value=0)" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "Nx8UebDLjCL-", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Print a few days\n", "print(activities_by_day_type.tail(12).to_string(formatters=formatters))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "zNOVoBwJRr3m", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Activities by day\n", "activities_by_day = activities_by_day_type.groupby('day').sum()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "Q_T5AAfURtmu", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Activities by type\n", "activities_by_type = activities_by_day_type.groupby('type').sum()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "nn6KeWl6RB2e", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Daily average\n", "print(\n", " activities_by_day[['distance', 'elapsed_time']]\n", " .mean()\n", " .to_frame('mean')\n", " .T\n", " .to_string(formatters=detailed_formatters))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "iy9HdX6zueNF", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Quick sanity check\n", "\n", "print(\"average distance per day: %.1fkm\" % (\n", " activities['distance'].sum() /\n", " (last_day - first_day)))\n", "print(\"average time per day: %.1fmin\" % (\n", " activities['elapsed_time'].sum() /\n", " (last_day - first_day)))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "V6BReJIW4Qmj", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Find days, months, quarters, years with extremes\n", "\n", "frequencies = ['D', 'M', 'Q', 'Y']\n", "\n", "for f in frequencies:\n", " extremes_by_type = (\n", " activities_by_day_type\n", " .unstack('type')\n", " .groupby(pd.Grouper(freq=f)).sum()\n", " .agg(['idxmax', 'max']))\n", " extremes_by_type.index = [f, 'max']\n", "\n", " print(\n", " extremes_by_type.T\n", " .unstack(level=0)\n", " .swaplevel(axis=1)\n", " .sort_index(axis=1))\n", " print('freq=%s\\n' % f)" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "92bmpDsVtlzq", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Weekdays with largest means\n", "weekday_abbr = 'Mon Tue Wed Thu Fri Sat Sun'.split()\n", "print(\n", " activities_by_day_type[['distance', 'elapsed_time', 'commute', 'count']]\n", " .unstack('type')\n", " .groupby(lambda d: weekday_abbr[d.weekday]).mean()\n", " .agg(['idxmax', 'max'])\n", " .T.unstack(level=0)\n", " .swaplevel(axis=1).sort_index(axis=1))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "Pf2M5xKB3yPS", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Months of the year with largest means\n", "months = 'Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec'.split()\n", "print(\n", " activities_by_day_type[['distance', 'elapsed_time', 'commute', 'count']]\n", " .unstack('type')\n", " .groupby(pd.Grouper(freq='M')).sum()\n", " .groupby(lambda m: months[m.month-1]).mean()\n", " .agg(['idxmax', 'max'])\n", " .T.unstack(level=0)\n", " .swaplevel(axis=1).sort_index(axis=1))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "E83KyhMcxit0", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Quarters of the year with largest means\n", "print(\n", " activities_by_day_type[['distance', 'elapsed_time', 'commute', 'count']]\n", " .unstack('type')\n", " .groupby(pd.Grouper(freq='Q')).sum()\n", " .groupby(lambda d: \"Q%d\" % d.quarter).mean()\n", " .agg(['idxmax', 'max'])\n", " .T.unstack(level=0)\n", " .swaplevel(axis=1).sort_index(axis=1))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "wPlf42imXnEY", "colab_type": "text" }, "cell_type": "markdown", "source": [ "# Plotting" ] }, { "metadata": { "id": "ZKGz_9yK0f7F", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot running activity, quarterly, summed\n", "import matplotlib.pyplot as plt\n", "from matplotlib.ticker import StrMethodFormatter\n", "\n", "fig, (ax1, ax2, ax3) = (\n", " plt.subplots(\n", " nrows=3, ncols=1,\n", " sharex=True, sharey=False,\n", " figsize=(8.025, 10)))\n", "\n", "runs_q_sum = (\n", " activities.loc[(slice(None), 'Run'), :]\n", " .reset_index('type', drop=True)\n", " .to_period('D')\n", " .groupby(pd.Grouper(freq='Q')).sum())\n", "\n", "runs_q_sum['distance'].plot(ax=ax1, kind='bar', color='#7799cc')\n", "ax1.set_ylabel('km')\n", "ax1.set_title('Distance')\n", "\n", "(runs_q_sum['elapsed_time'] / 60).plot(ax=ax2, kind='bar', color='#7799cc')\n", "ax2.set_ylabel('h')\n", "ax2.set_title('Duration')\n", "\n", "runs_q_sum['count'].plot(ax=ax3, kind='bar', color='#7799cc')\n", "ax3.set_ylabel('number')\n", "ax3.set_title('Count')\n", "\n", "fig.autofmt_xdate()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "fissOhZ9aOqF", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot running activity, quarterly, mean\n", "\n", "from matplotlib import ticker\n", "\n", "fig2, (ax4, ax5) = (\n", " plt.subplots(\n", " nrows=2, ncols=1,\n", " sharex=True, sharey=False,\n", " figsize=(8.025, 10/3*2)))\n", "\n", "(runs_q_sum['distance'] / runs_q_sum['count']).plot(ax=ax4, kind='bar', color='#7799cc');\n", "ax4.set_ylabel('km')\n", "ax4.set_title('Mean distance')\n", "ax4.yaxis.set_major_locator(ticker.MultipleLocator(5))\n", "\n", "(runs_q_sum['elapsed_time'] / runs_q_sum['count']).plot(ax=ax5, kind='bar', color='#7799cc');\n", "ax5.set_ylabel('min')\n", "ax5.set_title('Mean duration')\n", "ax5.yaxis.set_major_locator(ticker.MultipleLocator(60))\n", "\n", "fig2.autofmt_xdate()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "hruPUXTjcQHR", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot running activity, quarterly, median\n", "fig3, (ax6, ax7) = (\n", " plt.subplots(\n", " nrows=2, ncols=1,\n", " sharex=True, sharey=False,\n", " figsize=(8.025, 10/3*2)))\n", "\n", "runs_q_median = (\n", " activities.loc[(slice(None), 'Run'), :]\n", " .reset_index('type', drop=True)\n", " .to_period('D')\n", " .groupby(pd.Grouper(freq='Q')).median())\n", "\n", "runs_q_median['distance'].plot(ax=ax6, kind='bar', color='#7799cc')\n", "ax6.set_ylabel('km')\n", "ax6.set_title('Median distance')\n", "\n", "runs_q_median['elapsed_time'].plot(ax=ax7, kind='bar', color='#7799cc')\n", "ax7.set_ylabel('min')\n", "ax7.set_title('Median duration')\n", "ax7.yaxis.set_major_locator(ticker.MultipleLocator(30))\n", "\n", "fig3.autofmt_xdate()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "Lz5vBqBSgNTF", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot running activity, quarterly, pace\n", "fig4, (ax8, ax9) = (\n", " plt.subplots(\n", " nrows=2, ncols=1,\n", " sharex=True, sharey=False,\n", " figsize=(8.025, 10/3*2)))\n", "\n", "(runs_q_sum['elapsed_time'] / runs_q_sum['distance']).plot(ax=ax8, kind='bar', color='#7799cc');\n", "ax8.set_ylabel('min / km')\n", "ax8.set_title('Pace')\n", "\n", "(runs_q_sum['distance'] / (runs_q_sum['elapsed_time'] / 60)).plot(ax=ax9, kind='bar', color='#7799cc');\n", "ax9.set_ylabel('km / h')\n", "ax9.set_title('Speed')\n", "\n", "fig4.autofmt_xdate()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "pR73_kFXXjY9", "colab_type": "text" }, "cell_type": "markdown", "source": [ "# Machine learning" ] }, { "metadata": { "id": "YY7LQE7NZWW0", "colab_type": "text" }, "cell_type": "markdown", "source": [ "Strava allows users to tag activities that were commutes. It is very easy to forget to tag an activity, which then reduces the quality of the collected data. Here, we will build a classifier which predicts whether an activity was a commute, so that we can go back to Strava and add missing tags if necessary. Thus we will consider any activity, even those which happened after the activity for which we predict the commute label; _leakage_ (not-a-time-machine) should not be an issue here." ] }, { "metadata": { "id": "qEvSwsjnXdjn", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Split data into training and test\n", "\n", "import sklearn\n", "from sklearn.model_selection import train_test_split\n", "\n", "X = activities.reset_index()[['date', 'type', 'distance', 'elapsed_time', 'name']]\n", "y = activities.reset_index()['commute']\n", "\n", "X_train, X_test, y_train, y_test = (\n", " train_test_split(\n", " X, y,\n", " train_size=0.8,\n", " test_size=0.2,\n", " random_state=314,\n", " shuffle=True))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "wu7VG4V0MGnv", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Extract features\n", "\n", "import nltk\n", "from nltk.stem import WordNetLemmatizer\n", "\n", "nltk.download('wordnet')\n", "\n", "\n", "def type_features(X):\n", " return pd.get_dummies(X[['type']])\n", "\n", "\n", "def weekend_features(X):\n", " is_weekend = lambda d: d.dayofweek >= 5\n", " weekend = X['date'].apply(is_weekend)\n", " return pd.get_dummies(weekend, prefix='weekend')\n", "\n", "\n", "def time_features(X):\n", " time = pd.cut(X['date'].apply(lambda d: d.hour),\n", " right=False,\n", " bins=[0, 6, 9, 12, 15, 18, 21, 24],\n", " labels=['night',\n", " 'morning',\n", " 'late_morning',\n", " 'early_afternoon',\n", " 'late_afternoon',\n", " 'evening',\n", " 'late_evening'])\n", " return pd.get_dummies(time, prefix='time')\n", "\n", "\n", "def distance_features(X):\n", " distance = pd.cut(X['distance'],\n", " bins=np.logspace(1, 7, base=2, num=10),\n", " labels=['xxxs', 'xxs',\n", " 'xs', 's', 'm',\n", " 'l', 'xl', 'xxl', 'xxxl'])\n", " return pd.get_dummies(distance, prefix='distance')\n", "\n", "\n", "def duration_features(X):\n", " duration = pd.cut(X['elapsed_time'],\n", " bins=np.logspace(4, 8, base=2, num=10),\n", " labels=['xxxs', 'xxs',\n", " 'xs', 's', 'm',\n", " 'l', 'xl', 'xxl', 'xxxl'])\n", " return pd.get_dummies(duration, prefix='duration')\n", "\n", "\n", "def word_features(X, train_indices):\n", " lemmatizer = WordNetLemmatizer()\n", " lemmatize_words = lambda s: [lemmatizer.lemmatize(w) for w in s]\n", "\n", " word_separators = r'[,:()!? ]'\n", " word_hyphenators = r'[\\-_]'\n", " remove_short_words = lambda s: list(filter(lambda w: len(w) >= 5, s))\n", " make_unique = lambda s: list(set(s))\n", " prefix_with_word = lambda s: ['word_' + w for w in s]\n", "\n", " words = (\n", " X['name']\n", " .str.replace(word_separators, ' ')\n", " .str.replace(word_hyphenators, '')\n", " .str.lower()\n", " .str.split(' ')\n", " .apply(lambda s: [w.lower() for w in s])\n", " .apply(remove_short_words)\n", " .apply(make_unique)\n", " .apply(lemmatize_words)\n", " .apply(prefix_with_word)\n", " .apply(pd.Series)\n", " .stack()\n", " .reset_index(level=1, drop=True))\n", "\n", " words = words.reset_index().set_index(['index', 0])\n", " words['present'] = 1\n", " words = words.unstack(0, fill_value=0)\n", " words.columns.names = ['present', 'word']\n", " words.columns = words.columns.droplevel(level='present')\n", "\n", " # Correct for words only in test set; we could not possibly know\n", " # these words prior to prediction time.\n", " words_train = words.reindex(train_indices)\n", " words_only_in_test = (\n", " words_train.columns\n", " .to_series()\n", " .loc[words_train.sum() == 0]\n", " .index.values)\n", " words = words.drop(columns=words_only_in_test)\n", " return words\n", "\n", "\n", "def features(X_features_list, train_index, test_index):\n", " X_features = pd.concat(X_features_list, axis=1).fillna(value=0)\n", " return (\n", " X_features.iloc[train_index],\n", " X_features.iloc[test_index]\n", " )\n", "\n", "\n", "X_distance_duration_train, X_distance_duration_test = (\n", " features([\n", " distance_features(X),\n", " duration_features(X),\n", " ],\n", " X_train.index,\n", " X_test.index))\n", "\n", "X_no_word_train, X_no_word_test = (\n", " features([\n", " type_features(X),\n", " weekend_features(X),\n", " time_features(X),\n", " distance_features(X),\n", " duration_features(X),\n", " ],\n", " X_train.index,\n", " X_test.index))\n", "\n", "X_all_train, X_all_test = (\n", " features([\n", " type_features(X),\n", " weekend_features(X),\n", " time_features(X),\n", " distance_features(X),\n", " duration_features(X),\n", " word_features(X, X_train.index)\n", " ],\n", " X_train.index,\n", " X_test.index))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "b3qYPwsFqW3C", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Examine type features\n", "print(\n", " pd.crosstab(\n", " type_features(X_train).idxmax(axis=1),\n", " y_train,\n", " rownames=[''],\n", " colnames=['commute']))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "jPpQZGF4tF05", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Examine weekend features\n", "\n", "print(\n", " pd.crosstab(\n", " weekend_features(X_train).idxmax(axis=1),\n", " y_train,\n", " rownames=[''],\n", " colnames=['commute']))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "SH_HHp13qpPy", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Examine time features\n", "\n", "print(\n", " pd.crosstab(\n", " time_features(X_train).idxmax(axis=1),\n", " y_train,\n", " rownames=[''],\n", " colnames=['commute']))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "TCVQ6SfbtXv2", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Examine distance features\n", "\n", "print(\n", " pd.crosstab(\n", " distance_features(X_train).idxmax(axis=1),\n", " y_train,\n", " rownames=[''],\n", " colnames=['commute']))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "pnCz5IcdtXlL", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Examine duration features\n", "\n", "print(\n", " pd.crosstab(\n", " duration_features(X_train).idxmax(axis=1),\n", " y_train,\n", " rownames=[''],\n", " colnames=['commute']))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "oUNoPTCrF3DQ", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Establish baseline\n", "\n", "from sklearn.dummy import DummyClassifier\n", "\n", "dummy = DummyClassifier('most_frequent')\n", "dummy.fit(X_all_train, y_train)\n", "dummy.score(X_all_train, y_train)" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "uFxRiuPCG9h_", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Logistic Regression model\n", "\n", "from sklearn.linear_model import LogisticRegression\n", "from sklearn.model_selection import GridSearchCV\n", "\n", "\n", "def train_logistic_regression_model(X_features_train, y_train):\n", " lr_gs = GridSearchCV(LogisticRegression(random_state=99),\n", " {'C': [0.01, 0.1, 1.0, 10, 100]},\n", " scoring='roc_auc')\n", " lr_gs.fit(X_features_train, y_train)\n", " return lr_gs.best_estimator_, lr_gs.best_score_" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "7_HzItNCah47", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Naive Bayes classifier\n", "\n", "from sklearn.naive_bayes import BernoulliNB\n", "\n", "\n", "def train_naive_bayes_model(X_features_train, y_train):\n", " nb_gs = GridSearchCV(BernoulliNB(),\n", " {'alpha': [0.001, 0.01, 0.1, 1, 10]},\n", " scoring='roc_auc')\n", " nb_gs.fit(X_features_train, y_train)\n", " return nb_gs.best_estimator_, nb_gs.best_score_" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "PMSN_7xra-eG", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Support Vector Machine\n", "\n", "from sklearn.svm import SVC\n", "\n", "\n", "def train_support_vector_machine(X_features_train, y_train):\n", " svm_gs = GridSearchCV(SVC(probability=True, random_state=63),\n", " {'C': [0.01, 0.1, 1, 10, 100],\n", " 'gamma': [0.001, 0.01, 0.1, 1, 10]},\n", " scoring='roc_auc')\n", " svm_gs.fit(X_features_train, y_train)\n", " return svm_gs.best_estimator_, svm_gs.best_score_" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "nzi_YPLkRvyz", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title K-Nearest Neighbors\n", "\n", "from sklearn.neighbors import KNeighborsClassifier\n", "\n", "\n", "def train_knn_classifier(X_features_train, y_train):\n", " knn_gs = GridSearchCV(KNeighborsClassifier(p=1, weights='uniform'),\n", " {'n_neighbors': [1, 3, 5, 7, 9, 11, 13, 15]},\n", " scoring='roc_auc')\n", " knn_gs.fit(X_features_train, y_train)\n", " return knn_gs.best_estimator_, knn_gs.best_score_" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "Tvcff8SISZGJ", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Decision tree\n", "\n", "from sklearn.tree import DecisionTreeClassifier\n", "\n", "\n", "def train_decision_tree(X_features_train, y_train):\n", " tree_gs = GridSearchCV(DecisionTreeClassifier(splitter='best', random_state=112),\n", " {'min_samples_leaf': [5, 7, 9, 11, 13],\n", " 'max_depth': [None, 3, 5, 7, 9, 11]},\n", " scoring='roc_auc')\n", " tree_gs.fit(X_features_train, y_train)\n", " return tree_gs.best_estimator_, tree_gs.best_score_" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "qvVYWLsUGiE6", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Train models\n", "\n", "from sklearn.model_selection import cross_val_score\n", "\n", "training_functions = {\n", " 'Logistic Regression': train_logistic_regression_model,\n", " 'Naive Bayes': train_naive_bayes_model,\n", " 'Support Vector Machine': train_support_vector_machine,\n", " 'K-Nearest Neighbors': train_knn_classifier,\n", " 'Decision Tree': train_decision_tree,\n", "}\n", "\n", "features_train = {\n", " 'distance_duration': X_distance_duration_train,\n", " 'no_word': X_no_word_train,\n", " 'all': X_all_train,\n", "}\n", "\n", "models = []\n", "\n", "for c, f in training_functions.items():\n", " for d, x in features_train.items():\n", " m_n = '%s_%s' % (c, d)\n", " print('Training classifier \"%s\" with features \"%s\"...' % (c, d))\n", " m, _ = f(x, y_train)\n", " models.append({\n", " 'classifier': c,\n", " 'features': d,\n", " 'model': m,\n", " })" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "hKG7TsSfXOa3", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Test models\n", "\n", "from sklearn.metrics import f1_score\n", "from sklearn.metrics import precision_score\n", "from sklearn.metrics import recall_score\n", "from sklearn.metrics import roc_auc_score\n", "\n", "features_test = {\n", " 'distance_duration': X_distance_duration_test,\n", " 'no_word': X_no_word_test,\n", " 'all': X_all_test,\n", "}\n", "\n", "for m in models:\n", " X_features_train = features_train[m['features']]\n", " X_features_test = features_test[m['features']]\n", " y_score_train = m['model'].predict_proba(X_features_train)[:, 1]\n", " y_pred_train = m['model'].predict(X_features_train)\n", " y_pred_test = m['model'].predict(X_features_test)\n", " y_score_test = m['model'].predict_proba(X_features_test)[:, 1]\n", " m['AUC (training)'] = roc_auc_score(y_train, y_score_train, average='micro')\n", " m['AUC (test)'] = roc_auc_score(y_test, y_score_test, average='micro')\n", " m['F1 (training)'] = f1_score(y_train, y_pred_train, average='micro')\n", " m['F1 (test)'] = f1_score(y_test, y_pred_test, average='micro')\n", " m['Precision (training)'] = precision_score(y_train, y_pred_train, average='micro')\n", " m['Precision (test)'] = precision_score(y_test, y_pred_test, average='micro')\n", " m['Recall (training)'] = recall_score(y_train, y_pred_train, average='micro')\n", " m['Recall (test)'] = recall_score(y_test, y_pred_test, average='micro')\n", " \n", "models_df = pd.DataFrame.from_records(models, index=['classifier', 'features'])" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "iEpCw2hSlxbW", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot AUC, F1, Precision, Recall\n", "\n", "from matplotlib import ticker\n", "\n", "def plot_scores(ax, models_df, features, score, title):\n", " (models_df\n", " .loc[(slice(None), features), score]\n", " .reset_index('features')\n", " .sort_index()\n", " .plot(ax=ax, kind='barh', title=title, legend=True))\n", " ax.set_ylabel('')\n", " ax.set_xlim(0, 1)\n", " ax.xaxis.set_major_locator(ticker.FixedLocator(np.linspace(0, 1, 6)))\n", " \n", "def plot_all_scores(models_df, features):\n", " fig, ((ax1, ax2),\n", " (ax3, ax4)) = plt.subplots(nrows=2,\n", " ncols=2,\n", " sharex=True, \n", " sharey=True,\n", " figsize=(6, 6))\n", " fig.suptitle('Feature set: %s' % features, fontsize=11, color='#444444')\n", " plot_scores(ax1, models_df, features, ['AUC (test)', 'AUC (training)'], 'AUC')\n", " plot_scores(ax2, models_df, features, ['F1 (test)', 'F1 (training)'], 'F1')\n", " plot_scores(ax3, models_df, features, ['Precision (test)', 'Precision (training)'], 'Precision')\n", " plot_scores(ax4, models_df, features, ['Recall (test)', 'Recall (training)'], 'Recall')\n", "\n", " \n", "for f in features_test.keys():\n", " plot_all_scores(models_df, f)\n", "\n", "fig.tight_layout()" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "IxcYQehQXzZt", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Inspect coefficients from Logistic Regression model\n", "\n", "coef = pd.Series(models_df.loc[('Logistic Regression', 'no_word'), 'model'].coef_[0], index=X_no_word_train.columns)\n", "print('\\nPositive:')\n", "print(coef.loc[coef > 0].sort_values(ascending=False))\n", "print('\\nNegative:')\n", "print(coef.loc[coef < 0].sort_values(ascending=True))" ], "execution_count": 0, "outputs": [] }, { "metadata": { "id": "z03VlqJPS2bl", "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, "cellView": "form" }, "cell_type": "code", "source": [ "#@title Plot ROC curves\n", "from sklearn.metrics import roc_auc_score\n", "from sklearn.metrics import roc_curve\n", "\n", "def plot_roc_curve(ax, classifier, features):\n", " decision_function = (\n", " models_df.loc[(classifier, features), 'model']\n", " .decision_function(features_test[features]))\n", " fpr, tpr, _ = roc_curve(y_test, decision_function)\n", " ax.plot(fpr, tpr, color='darkorange', label='ROC curve')\n", " ax.fill_between(fpr, 0, tpr, color='#FFDFBA')\n", " ax.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')\n", " ax.set_xlabel('False Positive Rate')\n", " ax.set_ylabel('True Positive Rate')\n", " ax.set_title('%s' % features)\n", " ax.legend(loc=\"lower right\")\n", " \n", "def plot_roc_curves_for_logistic_regression_and_svm():\n", " fig, ((ax1, ax2),\n", " (ax3, ax4),\n", " (ax5, ax6)) = plt.subplots(nrows=3,\n", " ncols=2,\n", " sharex=True, \n", " sharey=True,\n", " figsize=(6, 9))\n", "\n", " plot_roc_curve(ax1, 'Logistic Regression', 'distance_duration')\n", " plot_roc_curve(ax2, 'Support Vector Machine', 'distance_duration')\n", " plot_roc_curve(ax3, 'Logistic Regression', 'no_word')\n", " plot_roc_curve(ax4, 'Support Vector Machine', 'no_word')\n", " plot_roc_curve(ax5, 'Logistic Regression', 'all')\n", " plot_roc_curve(ax6, 'Support Vector Machine', 'all')\n", "\n", " fig.tight_layout()\n", "\n", "plot_roc_curves_for_logistic_regression_and_svm()" ], "execution_count": 0, "outputs": [] } ] }