08_ensemble_hool.ipynb

{"metadata":{"orig_nbformat":4,"kernelspec":{"display_name":"Python 3 (ipykernel)","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":"3.7.12"}},"nbformat_minor":5,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Ensemble Learning mit dem Titanic-Datensatz\n\n- Der originale Datensatz kann hier abgerufen werden: https://www.kaggle.com/c/titanic\n- Der hier verwendete Datensatz wurde zwecks Einfachheit bereits überarbeitet und angereichtert","metadata":{},"id":"dff4f015-2dd7-4bce-8d42-3120c718d08a"},{"cell_type":"markdown","source":"## Lade Libraries und Daten","metadata":{},"id":"5d741c6c-9709-41de-bad2-16b4b3be8339"},{"cell_type":"code","source":"import pandas as pd\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport numpy as np\nimport re\n\nfrom sklearn.model_selection import train_test_split\n\npd.options.mode.chained_assignment = None\n\nDF = pd.read_csv('data/titanic_enh.csv')","metadata":{"trusted":true},"execution_count":1,"outputs":[],"id":"1374241d-2ce6-48ab-9da2-337c5e24cdf3"},{"cell_type":"code","source":"DF","metadata":{"trusted":true},"execution_count":2,"outputs":[{"execution_count":2,"output_type":"execute_result","data":{"text/plain":"     PassengerId  Survived  Pclass  \\\n0              1         0       3   \n1              2         1       1   \n2              3         1       3   \n3              4         1       1   \n4              5         0       3   \n..           ...       ...     ...   \n886          887         0       2   \n887          888         1       1   \n888          889         0       3   \n889          890         1       1   \n890          891         0       3   \n\n                                                  Name  Sex  Age  SibSp  \\\n0                              Braund, Mr. Owen Harris    1    1      1   \n1    Cumings, Mrs. John Bradley (Florence Briggs Th...    0    2      1   \n2                               Heikkinen, Miss. Laina    0    1      0   \n3         Futrelle, Mrs. Jacques Heath (Lily May Peel)    0    2      1   \n4                             Allen, Mr. William Henry    1    2      0   \n..                                                 ...  ...  ...    ...   \n886                              Montvila, Rev. Juozas    1    1      0   \n887                       Graham, Miss. Margaret Edith    0    1      0   \n888           Johnston, Miss. Catherine Helen \"Carrie\"    0    1      1   \n889                              Behr, Mr. Karl Howell    1    1      0   \n890                                Dooley, Mr. Patrick    1    1      0   \n\n     Parch            Ticket  Fare Cabin  Embarked  Has_Cabin  FamilySize  \\\n0        0         A/5 21171     0   NaN         0          0           2   \n1        0          PC 17599     3   C85         1          1           2   \n2        0  STON/O2. 3101282     1   NaN         0          0           1   \n3        0            113803     3  C123         0          1           2   \n4        0            373450     1   NaN         0          0           1   \n..     ...               ...   ...   ...       ...        ...         ...   \n886      0            211536     1   NaN         0          0           1   \n887      0            112053     2   B42         0          1           1   \n888      2        W./C. 6607     2   NaN         0          0           4   \n889      0            111369     2  C148         1          1           1   \n890      0            370376     0   NaN         2          0           1   \n\n     IsAlone  Title  \n0          0      1  \n1          0      3  \n2          1      4  \n3          0      3  \n4          1      1  \n..       ...    ...  \n886        1      5  \n887        1      4  \n888        0      4  \n889        1      1  \n890        1      1  \n\n[891 rows x 16 columns]","text/html":"<div>\n<style scoped>\n    .dataframe tbody tr th:only-of-type {\n        vertical-align: middle;\n    }\n\n    .dataframe tbody tr th {\n        vertical-align: top;\n    }\n\n    .dataframe thead th {\n        text-align: right;\n    }\n</style>\n<table border=\"1\" class=\"dataframe\">\n  <thead>\n    <tr style=\"text-align: right;\">\n      <th></th>\n      <th>PassengerId</th>\n      <th>Survived</th>\n      <th>Pclass</th>\n      <th>Name</th>\n      <th>Sex</th>\n      <th>Age</th>\n      <th>SibSp</th>\n      <th>Parch</th>\n      <th>Ticket</th>\n      <th>Fare</th>\n      <th>Cabin</th>\n      <th>Embarked</th>\n      <th>Has_Cabin</th>\n      <th>FamilySize</th>\n      <th>IsAlone</th>\n      <th>Title</th>\n    </tr>\n  </thead>\n  <tbody>\n    <tr>\n      <th>0</th>\n      <td>1</td>\n      <td>0</td>\n      <td>3</td>\n      <td>Braund, Mr. Owen Harris</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>A/5 21171</td>\n      <td>0</td>\n      <td>NaN</td>\n      <td>0</td>\n      <td>0</td>\n      <td>2</td>\n      <td>0</td>\n      <td>1</td>\n    </tr>\n    <tr>\n      <th>1</th>\n      <td>2</td>\n      <td>1</td>\n      <td>1</td>\n      <td>Cumings, Mrs. John Bradley (Florence Briggs Th...</td>\n      <td>0</td>\n      <td>2</td>\n      <td>1</td>\n      <td>0</td>\n      <td>PC 17599</td>\n      <td>3</td>\n      <td>C85</td>\n      <td>1</td>\n      <td>1</td>\n      <td>2</td>\n      <td>0</td>\n      <td>3</td>\n    </tr>\n    <tr>\n      <th>2</th>\n      <td>3</td>\n      <td>1</td>\n      <td>3</td>\n      <td>Heikkinen, Miss. Laina</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>STON/O2. 3101282</td>\n      <td>1</td>\n      <td>NaN</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>4</td>\n    </tr>\n    <tr>\n      <th>3</th>\n      <td>4</td>\n      <td>1</td>\n      <td>1</td>\n      <td>Futrelle, Mrs. Jacques Heath (Lily May Peel)</td>\n      <td>0</td>\n      <td>2</td>\n      <td>1</td>\n      <td>0</td>\n      <td>113803</td>\n      <td>3</td>\n      <td>C123</td>\n      <td>0</td>\n      <td>1</td>\n      <td>2</td>\n      <td>0</td>\n      <td>3</td>\n    </tr>\n    <tr>\n      <th>4</th>\n      <td>5</td>\n      <td>0</td>\n      <td>3</td>\n      <td>Allen, Mr. William Henry</td>\n      <td>1</td>\n      <td>2</td>\n      <td>0</td>\n      <td>0</td>\n      <td>373450</td>\n      <td>1</td>\n      <td>NaN</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n    </tr>\n    <tr>\n      <th>...</th>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n      <td>...</td>\n    </tr>\n    <tr>\n      <th>886</th>\n      <td>887</td>\n      <td>0</td>\n      <td>2</td>\n      <td>Montvila, Rev. Juozas</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>211536</td>\n      <td>1</td>\n      <td>NaN</td>\n      <td>0</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>5</td>\n    </tr>\n    <tr>\n      <th>887</th>\n      <td>888</td>\n      <td>1</td>\n      <td>1</td>\n      <td>Graham, Miss. Margaret Edith</td>\n      <td>0</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>112053</td>\n      <td>2</td>\n      <td>B42</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n      <td>4</td>\n    </tr>\n    <tr>\n      <th>888</th>\n      <td>889</td>\n      <td>0</td>\n      <td>3</td>\n      <td>Johnston, Miss. Catherine Helen \"Carrie\"</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>2</td>\n      <td>W./C. 6607</td>\n      <td>2</td>\n      <td>NaN</td>\n      <td>0</td>\n      <td>0</td>\n      <td>4</td>\n      <td>0</td>\n      <td>4</td>\n    </tr>\n    <tr>\n      <th>889</th>\n      <td>890</td>\n      <td>1</td>\n      <td>1</td>\n      <td>Behr, Mr. Karl Howell</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>111369</td>\n      <td>2</td>\n      <td>C148</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n    </tr>\n    <tr>\n      <th>890</th>\n      <td>891</td>\n      <td>0</td>\n      <td>3</td>\n      <td>Dooley, Mr. Patrick</td>\n      <td>1</td>\n      <td>1</td>\n      <td>0</td>\n      <td>0</td>\n      <td>370376</td>\n      <td>0</td>\n      <td>NaN</td>\n      <td>2</td>\n      <td>0</td>\n      <td>1</td>\n      <td>1</td>\n      <td>1</td>\n    </tr>\n  </tbody>\n</table>\n<p>891 rows × 16 columns</p>\n</div>"},"metadata":{}}],"id":"9fac9833-7b73-42dc-88ed-79bd1ee7f643"},{"cell_type":"markdown","source":"### Lade nötige Library/Libraries\n\nIn diesem Fall verwenden wir Decision Trees als Ausgangslage für das Ensemble Learning","metadata":{},"id":"19f796fd-5cf0-45f8-a5a8-bd536b588ffd"},{"cell_type":"code","source":"from sklearn import tree\nfrom sklearn.tree import DecisionTreeClassifier","metadata":{"trusted":true},"execution_count":3,"outputs":[],"id":"e3cdd861-68e0-434c-b835-2b46a987142e"},{"cell_type":"markdown","source":"### Trainiere drei weak learners parallel auf einem zufälligen Subset der Trainingsdaten (Bagging)","metadata":{},"id":"d5544c1b-3466-44c7-bcec-439a6d0f9a6a"},{"cell_type":"code","source":"# Sample for KFold\ndf = DF.copy()\ntrain, test = train_test_split(df, test_size=0.2)\n\n# Create Input- and Output-Structure and train Decision Tree Classifer\ntraining_columns = ['Pclass', 'Sex', 'Age','Fare','Embarked', 'Has_Cabin', 'FamilySize', 'IsAlone', 'Title']\n\n# Train three Decision Trees on subset of training data\nlearners = []\n\nfor i in range(3):\n    clf = DecisionTreeClassifier(max_depth = 3, criterion='entropy')\n    \n    sub_train, sub_test = train_test_split(train, test_size=0.8)\n    \n    X = sub_train[training_columns]\n    y = sub_train[['Survived']]\n    \n    clf = clf.fit(X, y)\n    learners.append(clf)","metadata":{"trusted":true},"execution_count":4,"outputs":[],"id":"efa4b08d-15b2-4c29-8498-6ae61954b565"},{"cell_type":"markdown","source":"### Treffe Vorhersagen mit den Weak Learners basierend auf einer Mehrheitswahl (Majority Vote)","metadata":{},"id":"3b690ac9-7937-4cf0-aaf8-fda40abbc853"},{"cell_type":"code","source":"# Create Predictions based on the predictions of the three weak learners\nX = test[training_columns]\ny = test[['Survived']]\n\nfor i in range(3):\n    y['Prediction_{}'.format(i)] = learners[i].predict(X)\n\n\ny['Prediction_Final'] = y.iloc[:, -3:].sum(axis=1)\ny['Prediction_Final'] = y['Prediction_Final'].apply(lambda x: 1 if x >= 3/2 else 0)\n\n# Calculate Performance Metric\ny['match'] = y['Survived'] == y['Prediction_Final']\nperformance =  sum(y['match']) / len(y)\nprint('Performance: {:.1%}'.format(performance))","metadata":{"trusted":true},"execution_count":5,"outputs":[{"name":"stdout","text":"Performance: 80.4%\n","output_type":"stream"}],"id":"1ce6d5e1-b635-439d-8045-fde451bdb170"},{"cell_type":"markdown","source":"### Suche Hyperparameter\n\nEs können mehrere Parameter gesucht werden (Anzahl Weak learners, Grösse der Trainingsdaten, Maximale Tiefe des Baums, Voting-Regel, ...) in diesem Fall suchen wir die optimale Grösse der Trainingsdaten mit drei weak learners.","metadata":{},"id":"489ccafc-00b8-4895-9af8-b2224645366b"},{"cell_type":"code","source":"share_of_training_data = []\ntest_performance = []\nnumber_of_learners = 3\ntries = 10\nmax_performance = 0\nmax_performance_share = 0\n\nfor i in range(1, 11):\n    share = i/10\n    overall_perf = 0\n    for j in range(tries):\n\n        # Sample for KFold\n        df = DF.copy()\n        train, test = train_test_split(df, test_size=0.2)\n        # Create Input- and Output-Structure and train Decision Tree Classifer\n        training_columns = ['Pclass', 'Sex', 'Age', 'Fare', 'Embarked', 'Has_Cabin', 'FamilySize', 'IsAlone', 'Title']\n       \n        # Train three Decision Trees on subset of training data\n        learners = []\n        \n        for i in range(number_of_learners):\n            clf = DecisionTreeClassifier(max_depth = 3, criterion='entropy')\n            sub_train, sub_test = train_test_split(train, test_size=0.8)\n            X = sub_train[training_columns]\n            y = sub_train[['Survived']]\n    \n            clf = clf.fit(X, y)\n            learners.append(clf)\n            \n        # Create Predictions based on the predictions of the three weak learners\n        X = test[training_columns]\n        y = test[['Survived']]\n\n        for i in range(number_of_learners):\n            y['Prediction_{}'.format(i)] = learners[i].predict(X)\n\n\n        y['Prediction_Final'] = y.iloc[:, -3:].sum(axis=1)\n        y['Prediction_Final'] = y['Prediction_Final'].apply(lambda x: 1 if x >= number_of_learners/2 else 0)\n\n        # Calculate Performance Metric\n        y['match'] = y['Survived'] == y['Prediction_Final']\n        performance =  sum(y['match']) / len(y)\n        overall_perf += performance\n            \n        \n        \n    share_of_training_data.append(share)\n    test_performance.append(overall_perf/tries)\n    if overall_perf/tries > max_performance:\n        max_performance = overall_perf/tries\n        max_performance_share = share\n\nprint('Die höchste Performance von {:.1%} wurde mit einem Anteil von {} Prozent am Trainingsset erreicht.'.format(max_performance, max_performance_share*100))\ng = sns.lineplot(x=share_of_training_data, y=test_performance)\ng.axes.set_ylim(0,1)\ng","metadata":{"trusted":true},"execution_count":6,"outputs":[{"name":"stdout","text":"Die höchste Performance von 82.5% wurde mit einem Anteil von 90.0 Prozent am Trainingsset erreicht.\n","output_type":"stream"},{"execution_count":6,"output_type":"execute_result","data":{"text/plain":"<AxesSubplot:>"},"metadata":{}},{"output_type":"display_data","data":{"text/plain":"<Figure size 432x288 with 1 Axes>","image/png":"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\n"},"metadata":{"needs_background":"light"}}],"id":"c54e0749-5d14-410d-ab7a-d615d5711def"}]}