Kmeans as Feature Engineering Form

import pandas as pd import numpy as np import matplotlib.pyplot as plt import seaborn as sns sns.set() %matplotlib inline print('Setup Completed ^^')

data = pd.read_csv('abalone.csv') data.head().T

data.shape

data.dtypes

data.Sex.value_counts(normalize=True)

# Number of Rings distributions : data.Rings.value_counts().sort_index()

# seaborn styles sns.set_context('notebook') sns.set_style('white') # custom colors male = sns.color_palette()[1] female = sns.color_palette()[4] infant = sns.color_palette()[8] # set bins for histogram bin_range = np.array(range(1, 31)) # plot histogram of quality counts for red and white wines ax = plt.axes() for color, plot_color in zip(['M', 'F', 'I'], [male, female, infant]): q_data = data.loc[data.Sex==color, 'Rings'] q_data.hist(bins=bin_range, alpha=0.5, ax=ax, color=plot_color, label=color) ax.legend() ax.set(xlabel='Number of Rings', ylabel='Frequency') # force tick labels to be in middle of region ax.set_xlim(3,30) ax.set_xticks(bin_range+0.5) ax.set_xticklabels(bin_range); ax.grid('on')

sns.set_style('white') sns.countplot(data=data, y='Sex')

sns.barplot(x=data['Sex'], y=data['Rings'])

float_columns = [x for x in data.columns if x not in ['Sex', 'Rings']] # correlation matrix : corr_mat = data[float_columns].corr() # Strip out the diagnoal values for next step: for x in range(len(float_columns)): corr_mat.iloc[x,x] = 0.0 corr_mat

sns.heatmap(data=corr_mat,annot=True)

# pairwise maximal correlations: corr_mat.abs().idxmax()

skew_columns = (data[float_columns].skew().sort_values(ascending=False)) skew_columns = skew_columns.loc[skew_columns > 0.75] skew_columns

for col in skew_columns.index.tolist(): data[col] = np.log1p(data[col])

from sklearn.preprocessing import StandardScaler sc = StandardScaler() data[float_columns] = sc.fit_transform(data[float_columns]) data.head()

sns.set_context('notebook') sns.pairplot(data[float_columns])

from sklearn.cluster import KMeans kmeans = KMeans(n_clusters=3, random_state=42) kmeans = kmeans.fit(data[float_columns])

data['kmeans'] = kmeans.predict(data[float_columns])

(data[['Sex', 'kmeans']].groupby(['kmeans','Sex']).size().to_frame().rename(columns={0: 'number'}))

# create and fit a range of models: km_list = list() for clust in range(1, 21): km = KMeans(n_clusters=clust, random_state=42) km = km.fit(data[float_columns]) km_list.append(pd.Series({'clusters': clust, 'inertia':km.inertia_, 'model':km}))

plot_data = (pd.concat(km_list, axis=1) .T [['clusters','inertia']] .set_index('clusters')) ax = plot_data.plot(marker='o', ls='-', color='purple') ax.set_xticks(range(0, 21, 2)) ax.set_xlim(0,21) ax.set(xlabel='Cluster', ylabel='Inertia');

from sklearn.cluster import AgglomerativeClustering ag = AgglomerativeClustering(n_clusters=3, linkage='ward', compute_full_tree=True) ag = ag.fit(data[float_columns]) data['agglom'] = ag.fit_predict(data[float_columns])

data[['Sex', 'agglom', 'kmeans']].groupby(['Sex', 'agglom']).size().to_frame().rename(columns={0:'number'})

(data[['Sex','agglom','kmeans']] .groupby(['Sex','agglom','kmeans']) .size() .to_frame() .rename(columns={0:'number'}))

data.describe(include='object').T

data.describe().T

from sklearn.ensemble import RandomForestClassifier from sklearn.metrics import classification_report, roc_auc_score from sklearn.model_selection import StratifiedShuffleSplit

y = (data['Rings'] > 9).astype(int) X_with_kmeans = data.drop(['agglom', 'Sex', 'Rings'], axis=1) X_without_kmeans = X_with_kmeans.drop('kmeans', axis=1)

sss = StratifiedShuffleSplit(n_splits=10, random_state=6535)

def get_avg_roc_10splits(estimator, X, y): roc_auc_list = [] for train_index, test_index in sss.split(X, y): X_train, X_test = X.loc[train_index], X.iloc[test_index] y_train, y_test = y.loc[train_index], y.iloc[test_index] estimator.fit(X_train, y_train) y_predictor = estimator.predict(X_test) y_scored = estimator.predict_proba(X_test)[:, 1] roc_auc_list.append(roc_auc_score(y_test, y_scored)) return np.mean(roc_auc_list) estimator = RandomForestClassifier() roc_with_kmeans = get_avg_roc_10splits(estimator, X_with_kmeans, y) roc_without_kmeans = get_avg_roc_10splits(estimator, X_without_kmeans, y) print("Without kmeans cluster as input to Random Forest, roc-auc is \"{0}\"".format(roc_without_kmeans)) print("Using kmeans cluster as input to Random Forest, roc-auc is \"{0}\"".format(roc_with_kmeans))