Clasificacion de Diagnostico de Cancer de Mama con Machine Learning

Este analisis utiliza el dataset Wisconsin Breast Cancer para construir modelos de clasificacion que permitan distinguir entre tumores benignos y malignos a partir de caracteristicas extraidas de imagenes digitalizadas de biopsias por aspiracion con aguja fina (FNA). Se exploran tecnicas de preprocesamiento, manejo de desbalance de clases y multiples algoritmos de clasificacion.

import seaborn as sns
from sklearn.decomposition import PCA, TruncatedSVD
import matplotlib.patches as mpatches
import time

# Librerias de clasificadores
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.naive_bayes import GaussianNB
import collections


# Otras Librerias
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from imblearn.pipeline import make_pipeline as imbalanced_make_pipeline
from imblearn.over_sampling import SMOTE
from imblearn.under_sampling import NearMiss
from imblearn.metrics import classification_report_imbalanced
from sklearn.metrics import precision_score, recall_score, f1_score, roc_auc_score, accuracy_score, classification_report
from collections import Counter
from sklearn.model_selection import KFold, StratifiedKFold
import warnings
warnings.filterwarnings("ignore")

import numpy as np # linear algebra
import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)
%matplotlib inline
import matplotlib.pyplot as plt
import scipy.stats as stats
import xgboost as xgb
from sklearn.model_selection import KFold
from IPython.display import HTML, display
from sklearn.manifold import TSNE
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler

df = pd.read_csv('data/data.csv')

1. Analisis Exploratorio de Datos (EDA)

Comenzamos con un analisis exploratorio para comprender la estructura del dataset, identificar valores faltantes, detectar registros duplicados y obtener una vision general de las distribuciones de las variables.

# Convertimos diagnosis de B/M a 0/1
df['diagnosis'] = df['diagnosis'].map({'B': 0, 'M': 1})
df['diagnosis'].value_counts()

diagnosis
0    357
1    212
Name: count, dtype: int64

# definimos "y" como la variable diagnostico 
#definimos a "X" como el resto de variables sin diagnostico, unnamed:32, id

y = df.diagnosis # M or B 
list_drp = ['Unnamed: 32','id','diagnosis']
X = df.drop(list_drp,axis = 1 )

#cols_with_missing = [col for col in train.columns if train[col].isnull().any()]
cols_with_missing = df.isnull().sum()
cols_with_missing = cols_with_missing[cols_with_missing>0]
cols_with_missing.sort_values(inplace=True)
fig, ax = plt.subplots(figsize=(7,6))  
width = 0.70 # the width of the bars 
ind = np.arange(len(cols_with_missing))  # the x locations for the groups
ax.barh(ind, cols_with_missing, width, color="blue")
ax.set_yticks(ind+width/2)
ax.set_yticklabels(cols_with_missing.index, minor=False)
plt.xlabel('Count')
plt.ylabel('Features') 

Text(0, 0.5, 'Features')

#vemos que unnamed esta vacía, por lo que la sacamos de la base de datos
lista=['Unnamed: 32','id']
df=df.drop(lista,axis = 1 )

# definimos "y" como la variable diagnostico 
#definimos a "X" como el resto de variables sin diagnostico, unnamed:32, id

y = df.diagnosis # M or B 
list_drp = ['diagnosis']
X = df.drop(list_drp,axis = 1 )





print('Los casos benignos son el', round(df['diagnosis'].value_counts()[0]/len(df) * 100,2), '% del dataset')
print('Los casos malignos son el', round(df['diagnosis'].value_counts()[1]/len(df) * 100,2), '% del dataset')
#graficamos el número de caos benignos y malignos, y de esta forma verificar si la clase es balanceada
ax = sns.countplot(y,label="Count")       # M = 212, B = 357
B, M = y.value_counts()
print('Número de benignos: ',B)
print('Número de malignos : ',M)
ax.set_ylabel('Número de pacientes')
bars = ax.patches
half = int(len(bars)/2)
left_bars = bars[:half]
right_bars = bars[half:]
for left, right in zip(left_bars, right_bars):
    height_l = left.get_height()
    height_r = right.get_height()
    total = height_l + height_r
    ax.text(left.get_x() + left.get_width()/2., height_l + 40, '{0:.0%}'.format(height_l/total), ha="center")
    ax.text(right.get_x() + right.get_width()/2., height_r + 40, '{0:.0%}'.format(height_r/total), ha="center")

Los casos benignos son el 62.74 % del dataset
Los casos malignos son el 37.26 % del dataset

Número de benignos:  357
Número de malignos :  212

# Vemos si existen registros duplicados
dups = X.duplicated()
# hacemos print de los registros duplicados si es que existen
print(dups.any())
print(X[dups])

False
Empty DataFrame
Columns: [radius_mean, texture_mean, perimeter_mean, area_mean, smoothness_mean, compactness_mean, concavity_mean, concave points_mean, symmetry_mean, fractal_dimension_mean, radius_se, texture_se, perimeter_se, area_se, smoothness_se, compactness_se, concavity_se, concave points_se, symmetry_se, fractal_dimension_se, radius_worst, texture_worst, perimeter_worst, area_worst, smoothness_worst, compactness_worst, concavity_worst, concave points_worst, symmetry_worst, fractal_dimension_worst]
Index: []

[0 rows x 30 columns]

#Hacemos un gráfico de las variables para ver las estadísticas descriptivas de cada una
#también nos permite ver que variables estan 
X.describe().T.style.bar(subset=['mean'], color='#205ff2')\
                            .background_gradient(subset=['std'], cmap='Reds')\
                            .background_gradient(subset=['50%'], cmap='coolwarm')

	count	mean	std	min	25%	50%	75%	max
radius_mean	569.000000	14.127292	3.524049	6.981000	11.700000	13.370000	15.780000	28.110000
texture_mean	569.000000	19.289649	4.301036	9.710000	16.170000	18.840000	21.800000	39.280000
perimeter_mean	569.000000	91.969033	24.298981	43.790000	75.170000	86.240000	104.100000	188.500000
area_mean	569.000000	654.889104	351.914129	143.500000	420.300000	551.100000	782.700000	2501.000000
smoothness_mean	569.000000	0.096360	0.014064	0.052630	0.086370	0.095870	0.105300	0.163400
compactness_mean	569.000000	0.104341	0.052813	0.019380	0.064920	0.092630	0.130400	0.345400
concavity_mean	569.000000	0.088799	0.079720	0.000000	0.029560	0.061540	0.130700	0.426800
concave points_mean	569.000000	0.048919	0.038803	0.000000	0.020310	0.033500	0.074000	0.201200
symmetry_mean	569.000000	0.181162	0.027414	0.106000	0.161900	0.179200	0.195700	0.304000
fractal_dimension_mean	569.000000	0.062798	0.007060	0.049960	0.057700	0.061540	0.066120	0.097440
radius_se	569.000000	0.405172	0.277313	0.111500	0.232400	0.324200	0.478900	2.873000
texture_se	569.000000	1.216853	0.551648	0.360200	0.833900	1.108000	1.474000	4.885000
perimeter_se	569.000000	2.866059	2.021855	0.757000	1.606000	2.287000	3.357000	21.980000
area_se	569.000000	40.337079	45.491006	6.802000	17.850000	24.530000	45.190000	542.200000
smoothness_se	569.000000	0.007041	0.003003	0.001713	0.005169	0.006380	0.008146	0.031130
compactness_se	569.000000	0.025478	0.017908	0.002252	0.013080	0.020450	0.032450	0.135400
concavity_se	569.000000	0.031894	0.030186	0.000000	0.015090	0.025890	0.042050	0.396000
concave points_se	569.000000	0.011796	0.006170	0.000000	0.007638	0.010930	0.014710	0.052790
symmetry_se	569.000000	0.020542	0.008266	0.007882	0.015160	0.018730	0.023480	0.078950
fractal_dimension_se	569.000000	0.003795	0.002646	0.000895	0.002248	0.003187	0.004558	0.029840
radius_worst	569.000000	16.269190	4.833242	7.930000	13.010000	14.970000	18.790000	36.040000
texture_worst	569.000000	25.677223	6.146258	12.020000	21.080000	25.410000	29.720000	49.540000
perimeter_worst	569.000000	107.261213	33.602542	50.410000	84.110000	97.660000	125.400000	251.200000
area_worst	569.000000	880.583128	569.356993	185.200000	515.300000	686.500000	1084.000000	4254.000000
smoothness_worst	569.000000	0.132369	0.022832	0.071170	0.116600	0.131300	0.146000	0.222600
compactness_worst	569.000000	0.254265	0.157336	0.027290	0.147200	0.211900	0.339100	1.058000
concavity_worst	569.000000	0.272188	0.208624	0.000000	0.114500	0.226700	0.382900	1.252000
concave points_worst	569.000000	0.114606	0.065732	0.000000	0.064930	0.099930	0.161400	0.291000
symmetry_worst	569.000000	0.290076	0.061867	0.156500	0.250400	0.282200	0.317900	0.663800
fractal_dimension_worst	569.000000	0.083946	0.018061	0.055040	0.071460	0.080040	0.092080	0.207500

Mapa de Calor de Correlaciones

El mapa de calor nos permite visualizar la matriz de correlacion entre todas las variables continuas. Los colores representan la magnitud y direccion de la correlacion: valores cercanos a 1 (o -1) indican alta correlacion positiva (o negativa), mientras que valores cercanos a 0 sugieren poca relacion lineal.

corrmat = X.corr()
f, ax = plt.subplots(figsize=(12, 9))
sns.heatmap(corrmat, annot=True, linewidths=.5, fmt= '.1f',ax=ax);

Interpretacion de Correlaciones y Seleccion de Variables

El mapa de calor revela grupos de variables altamente correlacionadas entre si. Por ejemplo, radius_mean, perimeter_mean y area_mean presentan alta correlacion mutua, lo cual es esperable dado que las tres describen el tamano del nucleo celular. De manera similar, compactness_mean, concavity_mean y concave points_mean estan fuertemente correlacionadas.

Para reducir la multicolinealidad y mejorar la estabilidad de los modelos, seleccionamos un representante de cada grupo de variables correlacionadas, eliminando las redundantes.

drop_list1 = ['perimeter_mean','radius_mean','compactness_mean','concave points_mean','radius_se','perimeter_se','radius_worst','perimeter_worst','compactness_worst','concave points_worst','compactness_se','concave points_se','texture_worst','area_worst']
x_1 = X.drop(drop_list1,axis = 1 ) 
#correlation map
f,ax = plt.subplots(figsize=(14, 14))
sns.heatmap(x_1.corr(), annot=True, linewidths=.5, fmt= '.1f',ax=ax)

<Axes: >

2. Preprocesamiento

Aplicamos un escalado MinMaxScaler para normalizar las variables al rango [0, 1]. Esto es particularmente importante para algoritmos sensibles a la escala de las variables, como SVM. Ademas, preparamos los conjuntos de entrenamiento y prueba.

# visualize a minmax scaler transform of the sonar dataset
from pandas import read_csv
from pandas import DataFrame
from pandas.plotting import scatter_matrix
from sklearn.preprocessing import MinMaxScaler
from matplotlib import pyplot

data = df.values[:, :]
# perform a robust scaler transform of the dataset
trans = MinMaxScaler()
data = trans.fit_transform(data)
# convert the array back to a dataframe
dfbalanceado = DataFrame(data)
dfbalanceado.columns = df.columns
# summarize
print(dfbalanceado.describe())
# histograms of the variables
dfbalanceado.hist()
pyplot.show()

        diagnosis  radius_mean  texture_mean  perimeter_mean   area_mean  \
count  569.000000   569.000000    569.000000      569.000000  569.000000   
mean     0.372583     0.338222      0.323965        0.332935    0.216920   
std      0.483918     0.166787      0.145453        0.167915    0.149274   
min      0.000000     0.000000      0.000000        0.000000    0.000000   
25%      0.000000     0.223342      0.218465        0.216847    0.117413   
50%      0.000000     0.302381      0.308759        0.293345    0.172895   
75%      1.000000     0.416442      0.408860        0.416765    0.271135   
max      1.000000     1.000000      1.000000        1.000000    1.000000   

       smoothness_mean  compactness_mean  concavity_mean  concave points_mean  \
count       569.000000        569.000000      569.000000           569.000000   
mean          0.394785          0.260601        0.208058             0.243137   
std           0.126967          0.161992        0.186785             0.192857   
min           0.000000          0.000000        0.000000             0.000000   
25%           0.304595          0.139685        0.069260             0.100944   
50%           0.390358          0.224679        0.144189             0.166501   
75%           0.475490          0.340531        0.306232             0.367793   
max           1.000000          1.000000        1.000000             1.000000   

       symmetry_mean  ...  radius_worst  texture_worst  perimeter_worst  \
count     569.000000  ...    569.000000     569.000000       569.000000   
mean        0.379605  ...      0.296663       0.363998         0.283138   
std         0.138456  ...      0.171940       0.163813         0.167352   
min         0.000000  ...      0.000000       0.000000         0.000000   
25%         0.282323  ...      0.180719       0.241471         0.167837   
50%         0.369697  ...      0.250445       0.356876         0.235320   
75%         0.453030  ...      0.386339       0.471748         0.373475   
max         1.000000  ...      1.000000       1.000000         1.000000   

       area_worst  smoothness_worst  compactness_worst  concavity_worst  \
count  569.000000        569.000000         569.000000       569.000000   
mean     0.170906          0.404138           0.220212         0.217403   
std      0.139932          0.150779           0.152649         0.166633   
min      0.000000          0.000000           0.000000         0.000000   
25%      0.081130          0.300007           0.116337         0.091454   
50%      0.123206          0.397081           0.179110         0.181070   
75%      0.220901          0.494156           0.302520         0.305831   
max      1.000000          1.000000           1.000000         1.000000   

       concave points_worst  symmetry_worst  fractal_dimension_worst  
count            569.000000      569.000000               569.000000  
mean               0.393836        0.263307                 0.189596  
std                0.225884        0.121954                 0.118466  
min                0.000000        0.000000                 0.000000  
25%                0.223127        0.185098                 0.107700  
50%                0.343402        0.247782                 0.163977  
75%                0.554639        0.318155                 0.242949  
max                1.000000        1.000000                 1.000000  

[8 rows x 31 columns]

#Este código colocamos la variable diagnosis en la última columna por comodidad
diagnosis = dfbalanceado['diagnosis']

dfbalanceado.drop(['diagnosis'], axis=1, inplace=True)
dfbalanceado.insert(30, 'diagnosis', diagnosis)


dfbalanceado.head()

	radius_mean	texture_mean	perimeter_mean	area_mean	smoothness_mean	compactness_mean	concavity_mean	concave points_mean	symmetry_mean	fractal_dimension_mean	...	texture_worst	perimeter_worst	area_worst	smoothness_worst	compactness_worst	concavity_worst	concave points_worst	symmetry_worst	fractal_dimension_worst	diagnosis
0	0.521037	0.022658	0.545989	0.363733	0.593753	0.792037	0.703140	0.731113	0.686364	0.605518	...	0.141525	0.668310	0.450698	0.601136	0.619292	0.568610	0.912027	0.598462	0.418864	1.0
1	0.643144	0.272574	0.615783	0.501591	0.289880	0.181768	0.203608	0.348757	0.379798	0.141323	...	0.303571	0.539818	0.435214	0.347553	0.154563	0.192971	0.639175	0.233590	0.222878	1.0
2	0.601496	0.390260	0.595743	0.449417	0.514309	0.431017	0.462512	0.635686	0.509596	0.211247	...	0.360075	0.508442	0.374508	0.483590	0.385375	0.359744	0.835052	0.403706	0.213433	1.0
3	0.210090	0.360839	0.233501	0.102906	0.811321	0.811361	0.565604	0.522863	0.776263	1.000000	...	0.385928	0.241347	0.094008	0.915472	0.814012	0.548642	0.884880	1.000000	0.773711	1.0
4	0.629893	0.156578	0.630986	0.489290	0.430351	0.347893	0.463918	0.518390	0.378283	0.186816	...	0.123934	0.506948	0.341575	0.437364	0.172415	0.319489	0.558419	0.157500	0.142595	1.0

5 rows × 31 columns

print(dfbalanceado.columns
      )

Index(['radius_mean', 'texture_mean', 'perimeter_mean', 'area_mean',
       'smoothness_mean', 'compactness_mean', 'concavity_mean',
       'concave points_mean', 'symmetry_mean', 'fractal_dimension_mean',
       'radius_se', 'texture_se', 'perimeter_se', 'area_se', 'smoothness_se',
       'compactness_se', 'concavity_se', 'concave points_se', 'symmetry_se',
       'fractal_dimension_se', 'radius_worst', 'texture_worst',
       'perimeter_worst', 'area_worst', 'smoothness_worst',
       'compactness_worst', 'concavity_worst', 'concave points_worst',
       'symmetry_worst', 'fractal_dimension_worst', 'diagnosis'],
      dtype='str')

print(len(dfbalanceado.columns))
print(len(dfbalanceado
          ))

31
569

3. Manejo de Desbalance de Clases

El dataset presenta un desbalance entre las clases benigno y maligno. Para abordar este problema, evaluamos dos estrategias de remuestreo: sub-muestreo aleatorio (under-sampling) y sobre-muestreo aleatorio (over-sampling).

print('Benigno', round(dfbalanceado['diagnosis'].value_counts()[0]/len(dfbalanceado) * 100,2), '% del dataset')
print('Maligno', round(dfbalanceado['diagnosis'].value_counts()[1]/len(dfbalanceado) * 100,2), '% del dataset')

X = dfbalanceado.drop('diagnosis', axis=1)
y = dfbalanceado['diagnosis']

Xtrain, Xtest, ytrain, ytest = train_test_split(X, y, test_size=0.2, random_state=42)

#convirtiendo los datos en array
Xtrain = Xtrain.values
Xtest = Xtest.values
ytrain = ytrain.values
ytest = ytest.values

# Viendo si la distribución de las etiqueta de train y test se distribuyen de manera similar
train_unique_label, train_counts_label = np.unique(ytrain, return_counts=True)
test_unique_label, test_counts_label = np.unique(ytest, return_counts=True)
print('-' * 100)

print('Distribución de las etiquetas: \n')
print(train_counts_label/ len(ytrain))
print(test_counts_label/ len(ytest))
print(len(Xtrain))
print(len(Xtest))
print(len(ytrain))
print(len(ytest))
print(len(dfbalanceado))
print(len(dfbalanceado.columns))
print(len(X))
print(len(y))

Benigno 62.74 % del dataset
Maligno 37.26 % del dataset
----------------------------------------------------------------------------------------------------
Distribución de las etiquetas: 

[0.62857143 0.37142857]
[0.62280702 0.37719298]
455
114
455
114
569
31
569
569

3.1 Random Under-Sampling

El sub-muestreo aleatorio reduce la clase mayoritaria al tamano de la clase minoritaria, eliminando observaciones de forma aleatoria. Esto equilibra las clases pero puede descartar informacion util.

from imblearn.under_sampling import RandomUnderSampler

# Vamos a mezclar los datos antes de crear las submuestras.

rus = RandomUnderSampler(random_state=0)
rus.fit(Xtrain, ytrain)
X_train_undersampling, y_train_undersampling = rus.fit_resample(Xtrain, ytrain)

df_under_sampling = pd.DataFrame(X_train_undersampling, columns=dfbalanceado.columns[:-1])
df_under_sampling['diagnosis'] = y_train_undersampling
df_under_sampling

	radius_mean	texture_mean	perimeter_mean	area_mean	smoothness_mean	compactness_mean	concavity_mean	concave points_mean	symmetry_mean	fractal_dimension_mean	...	texture_worst	perimeter_worst	area_worst	smoothness_worst	compactness_worst	concavity_worst	concave points_worst	symmetry_worst	fractal_dimension_worst	diagnosis
0	0.308060	0.425769	0.297975	0.177094	0.314977	0.176676	0.111317	0.168191	0.378283	0.152064	...	0.527719	0.241994	0.126229	0.297365	0.139525	0.182268	0.440550	0.257441	0.092680	0.0
1	0.135643	0.201894	0.132748	0.063499	0.381782	0.198791	0.054592	0.120080	0.165152	0.399115	...	0.292377	0.119080	0.047016	0.467080	0.191140	0.067364	0.224330	0.184703	0.243015	0.0
2	0.239907	0.166385	0.236680	0.129714	0.455629	0.219434	0.154452	0.136630	0.310606	0.220514	...	0.231343	0.196574	0.097670	0.516608	0.182699	0.243610	0.225017	0.232998	0.183458	0.0
3	0.247953	0.349341	0.246562	0.131326	0.514309	0.293908	0.191542	0.107654	0.537374	0.399747	...	0.323827	0.172917	0.081130	0.455854	0.198126	0.282348	0.277938	0.225508	0.218746	0.0
4	0.204411	0.286777	0.208279	0.104305	0.390810	0.346973	0.362699	0.141849	0.502020	0.562974	...	0.424840	0.183027	0.070709	0.419534	0.443878	0.593930	0.418557	0.343584	0.489702	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
333	0.569786	0.503213	0.540460	0.395546	0.339984	0.310472	0.343955	0.411083	0.451010	0.133319	...	0.490139	0.510434	0.353372	0.354817	0.284571	0.459665	0.672165	0.471319	0.248196	1.0
334	0.428274	0.196145	0.428512	0.275589	0.381692	0.361082	0.282099	0.349950	0.364646	0.206403	...	0.265458	0.367996	0.217460	0.477646	0.407981	0.395847	0.680756	0.286615	0.237439	1.0
335	0.341190	0.476835	0.339161	0.198176	0.379164	0.341145	0.261246	0.321173	0.593434	0.302654	...	0.608475	0.321679	0.153878	0.559532	0.367329	0.299042	0.608935	0.622708	0.311951	1.0
336	0.552747	0.250592	0.536314	0.395970	0.476393	0.277958	0.341378	0.430666	0.457576	0.256318	...	0.343284	0.473081	0.335185	0.522552	0.195797	0.261342	0.575258	0.261975	0.193625	1.0
337	0.331251	0.335137	0.327068	0.193425	0.481809	0.288080	0.263824	0.321223	0.307576	0.326032	...	0.500533	0.316201	0.168133	0.595192	0.319692	0.325000	0.627835	0.318155	0.330972	1.0

338 rows × 31 columns

3.2 Random Over-Sampling

El sobre-muestreo aleatorio duplica observaciones de la clase minoritaria hasta igualar la clase mayoritaria. Esto preserva toda la informacion disponible, aunque puede introducir sobreajuste.

from imblearn.over_sampling import RandomOverSampler

ros = RandomOverSampler(random_state=0)
ros.fit(Xtrain, ytrain)
X_train_oversampling, y_train_oversampling = ros.fit_resample(Xtrain, ytrain)

df_over_sampling = pd.DataFrame(X_train_oversampling, columns=dfbalanceado.columns[:-1])
df_over_sampling['diagnosis'] = y_train_oversampling
df_over_sampling

	radius_mean	texture_mean	perimeter_mean	area_mean	smoothness_mean	compactness_mean	concavity_mean	concave points_mean	symmetry_mean	fractal_dimension_mean	...	texture_worst	perimeter_worst	area_worst	smoothness_worst	compactness_worst	concavity_worst	concave points_worst	symmetry_worst	fractal_dimension_worst	diagnosis
0	0.096928	0.257694	0.103656	0.045387	0.487226	0.373965	0.733365	0.217445	0.530808	0.642376	...	0.283316	0.075153	0.034285	0.508684	0.397018	1.000000	0.601375	0.524936	0.409681	0.0
1	0.667755	0.570172	0.683505	0.495228	0.554934	0.809214	0.582709	0.743539	0.674242	0.505897	...	0.571962	0.627970	0.467902	0.514627	0.709327	0.541534	0.997595	0.499310	0.481175	1.0
2	0.103744	0.140345	0.106489	0.049799	0.221901	0.208975	0.140300	0.108350	0.646970	0.414280	...	0.192164	0.075601	0.030697	0.179555	0.136324	0.111581	0.174811	0.338459	0.195855	0.0
3	0.173648	0.524518	0.167369	0.086320	0.396678	0.162444	0.055740	0.080268	0.422727	0.280750	...	0.617537	0.137308	0.066482	0.519910	0.109158	0.089856	0.210859	0.363493	0.173357	0.0
4	0.150930	0.174839	0.143459	0.071432	0.548614	0.187811	0.025398	0.064115	0.850000	0.413648	...	0.144723	0.096867	0.045075	0.371987	0.069244	0.017316	0.088625	0.392667	0.165027	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
567	0.550381	0.356442	0.541151	0.403181	0.377088	0.267530	0.349110	0.384245	0.321717	0.148062	...	0.406183	0.445690	0.299302	0.413590	0.178916	0.275240	0.512027	0.152967	0.125738	1.0
568	0.593923	0.769699	0.581922	0.457900	0.285005	0.287160	0.268276	0.329871	0.185859	0.066765	...	0.889925	0.646397	0.563262	0.459816	0.371016	0.319089	0.558419	0.226296	0.135380	1.0
569	0.404610	0.806561	0.414000	0.255101	0.484517	0.443286	0.410262	0.417445	0.520707	0.348357	...	1.000000	0.377957	0.208858	0.773493	0.513345	0.455511	0.692096	0.383797	0.428703	1.0
570	0.638885	0.397362	0.613019	0.493107	0.279137	0.196614	0.211856	0.299304	0.205556	0.038121	...	0.377132	0.554261	0.384585	0.340950	0.197737	0.252236	0.496564	0.132663	0.106454	1.0
571	0.397037	0.441326	0.389814	0.248017	0.355421	0.258328	0.262887	0.371918	0.331818	0.231887	...	0.368337	0.284327	0.158695	0.360100	0.167273	0.227316	0.507216	0.195348	0.086842	1.0

572 rows × 31 columns

colors = ["#0101DF", "#DF0101"]
sns.countplot(x='diagnosis', data=df_under_sampling, palette=colors)
plt.title('Clases igualmente distribuidas', fontsize=14)
plt.show()

3.3 Matrices de Correlacion Post-Remuestreo

Comparamos las matrices de correlacion del dataset completo, el sub-muestreado y el sobre-muestreado para verificar que las relaciones entre variables se mantienen consistentes tras el remuestreo.

# Hay que asegurarse de usar la submuestra en nuestra correlación
f, (ax1, ax2, ax3) = plt.subplots(3, 1, figsize=(24,20))

# DataFrame completo
corr = dfbalanceado.corr()
sns.heatmap(corr, cmap='coolwarm_r', annot_kws={'size':20}, ax=ax1)
ax1.set_title("Matriz de correlación con imbalance \n (no se usa como referencia)", fontsize=14)

sub_sample_corr = df_under_sampling.corr()
sns.heatmap(sub_sample_corr, cmap='coolwarm_r', annot_kws={'size':20}, ax=ax2)
ax2.set_title('Matriz de correlación usando submuestra \n (usar como referencia)', fontsize=14)

over_sample_corr = df_over_sampling.corr()
sns.heatmap(over_sample_corr, cmap='coolwarm_r', annot_kws={'size':20}, ax=ax3)
ax3.set_title('Matriz de correlación usando sobremuestra \n (usar como referencia)', fontsize=14)
plt.show()

4. Modelamiento y Evaluacion (Under-Sampling)

Entrenamos tres clasificadores sobre el conjunto sub-muestreado: Support Vector Classifier (SVC), Decision Tree y Naive Bayes. Para cada uno, realizamos validacion cruzada y optimizacion de hiperparametros mediante GridSearchCV.

# Se obtiene X e y a partir del nuevo dataset
X_train_undersampling = df_under_sampling.drop('diagnosis', axis=1)
y_train_undersampling = df_under_sampling['diagnosis']

classifiers = {
    "Support Vector Classifier": SVC(),
    "DecisionTreeClassifier": DecisionTreeClassifier(),
    "Naive_bayes": GaussianNB()
}

from sklearn.model_selection import cross_val_score

for key, classifier in classifiers.items():
    classifier.fit(X_train_undersampling, y_train_undersampling)
    training_score = cross_val_score(classifier, X_train_undersampling, y_train_undersampling, cv=5)
    print("Clasificador: ", classifier.__class__.__name__, "tiene una precisión del", round(training_score.mean(), 2) * 100, "%")

Clasificador:  SVC tiene una precisión del 97.0 %
Clasificador:  DecisionTreeClassifier tiene una precisión del 90.0 %
Clasificador:  GaussianNB tiene una precisión del 92.0 %

#GridSearchCV para encontrar los mejores hiperparámetros

from sklearn.model_selection import GridSearchCV

# Support Vector Classifier
svc_params = {'C': [0.5, 0.7, 0.9, 1], 'kernel': ['rbf', 'poly', 'sigmoid', 'linear']}
grid_svc = GridSearchCV(SVC(), svc_params)
grid_svc.fit(X_train_undersampling, y_train_undersampling)

# mejor modelo SVC  
svc = grid_svc.best_estimator_

# DecisionTree 
tree_params = {"criterion": ["gini", "entropy"], "max_depth": list(range(2,4,1)), 
              "min_samples_leaf": list(range(5,7,1))}
grid_tree = GridSearchCV(DecisionTreeClassifier(), tree_params)
grid_tree.fit(X_train_undersampling, y_train_undersampling)

# mejor modelo árbol de decisión  
tree_clf = grid_tree.best_estimator_

# Naive 
naive_clf = GaussianNB()
naive_clf.fit(X_train_undersampling, y_train_undersampling)

GaussianNB()

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

svc

SVC(C=0.7, kernel='linear')

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

tree_clf

DecisionTreeClassifier(max_depth=2, min_samples_leaf=6)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

svc_score = cross_val_score(svc, X_train_undersampling, y_train_undersampling, cv=5)
print('Support Vector Classifier Cross Validation Score', round(svc_score.mean() * 100, 2).astype(str) + '%')

tree_score = cross_val_score(tree_clf, X_train_undersampling, y_train_undersampling, cv=5)
print('DecisionTree Classifier Cross Validation Score', round(tree_score.mean() * 100, 2).astype(str) + '%')

naive_score = cross_val_score(naive_clf, X_train_undersampling, y_train_undersampling, cv=5)
print('Naive Bayes Classifier Cross Validation Score', round(naive_score.mean() * 100, 2).astype(str) + '%')

Support Vector Classifier Cross Validation Score 97.63%
DecisionTree Classifier Cross Validation Score 92.58%
Naive Bayes Classifier Cross Validation Score 91.99%

from sklearn.metrics import roc_curve
from sklearn.model_selection import cross_val_predict
# Crea un DataFrame con todos los puntajes y los nombres de los clasificadores.

svc_pred = cross_val_predict(svc, X_train_undersampling, y_train_undersampling, cv=5,
                             method="decision_function")

tree_pred = cross_val_predict(tree_clf, X_train_undersampling, y_train_undersampling, cv=5)

naive_pred = cross_val_predict(naive_clf, X_train_undersampling, y_train_undersampling, cv=5)

from sklearn.metrics import roc_auc_score

print('Support Vector Classifier: ', roc_auc_score(y_train_undersampling, svc_pred))
print('Decision Tree Classifier: ', roc_auc_score(y_train_undersampling, tree_pred))
print('Naive Bayes Tree Classifier: ', roc_auc_score(y_train_undersampling, naive_pred))

Support Vector Classifier:  0.992297188473793
Decision Tree Classifier:  0.9289940828402368
Naive Bayes Tree Classifier:  0.9201183431952662

4.1 Matriz de Confusion y Metricas (Under-Sampling)

Evaluamos el desempeno de los modelos entrenados con under-sampling sobre el conjunto de prueba. La matriz de confusion nos permite visualizar los verdaderos positivos, falsos positivos, verdaderos negativos y falsos negativos de cada clasificador.

df_test = pd.DataFrame(Xtest, columns=dfbalanceado.columns[:-1])
df_test['diagnosis'] = ytest


print('Distribución de las clases en el conjunto de datos de submuestra - OVERSAMPLING')
print(df_test['diagnosis'].value_counts())



sns.countplot(x='diagnosis', data=df_test, palette=colors)
plt.title('Clases igualmente distribuidas', fontsize=14)
plt.show()

Distribución de las clases en el conjunto de datos de submuestra - OVERSAMPLING
diagnosis
0.0    71
1.0    43
Name: count, dtype: int64

from sklearn.metrics import confusion_matrix

y_pred_svc = svc.predict(Xtest)
y_pred_tree = tree_clf.predict(Xtest)
y_pred_naive = naive_clf.predict(Xtest)


svc_cf = confusion_matrix(ytest, y_pred_svc)
tree_cf = confusion_matrix(ytest, y_pred_tree)
naive_cf = confusion_matrix(ytest, y_pred_naive)

fig, ax = plt.subplots(2, 2,figsize=(22,12))


sns.heatmap(naive_cf, ax=ax[0][1], annot=True, cmap=plt.cm.copper)
ax[0][1].set_title("Naive Bayes \n Matriz de Confusión", fontsize=14)
ax[0][1].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[0][1].set_yticklabels(['', ''], fontsize=14, rotation=360)

sns.heatmap(svc_cf, ax=ax[1][0], annot=True, cmap=plt.cm.copper)
ax[1][0].set_title("Suppor Vector Machine \n Matriz de Confusión", fontsize=14)
ax[1][0].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[1][0].set_yticklabels(['', ''], fontsize=14, rotation=360)

sns.heatmap(tree_cf, ax=ax[1][1], annot=True, cmap=plt.cm.copper)
ax[1][1].set_title("DecisionTree \n Matriz de Confusión", fontsize=14)
ax[1][1].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[1][1].set_yticklabels(['', ''], fontsize=14, rotation=360)


plt.show()

y_true = [0, 1, 2, 2, 2]
y_pred = [0, 0, 2, 2, 1]
target_names = ['Benigno', 'Maligno']
print("Reporte clasificación SVC")
print(classification_report(ytest, y_pred_svc, target_names=target_names))
print("")
print("Reporte clasificación Árbol")
print(classification_report(ytest, y_pred_tree, target_names=target_names))
print("")
print("Reporte clasificación Naive Bayes")
print(classification_report(ytest, y_pred_naive, target_names=target_names))

Reporte clasificación SVC
              precision    recall  f1-score   support

     Benigno       0.97      0.99      0.98        71
     Maligno       0.98      0.95      0.96        43

    accuracy                           0.97       114
   macro avg       0.97      0.97      0.97       114
weighted avg       0.97      0.97      0.97       114


Reporte clasificación Árbol
              precision    recall  f1-score   support

     Benigno       0.97      0.94      0.96        71
     Maligno       0.91      0.95      0.93        43

    accuracy                           0.95       114
   macro avg       0.94      0.95      0.94       114
weighted avg       0.95      0.95      0.95       114


Reporte clasificación Naive Bayes
              precision    recall  f1-score   support

     Benigno       0.96      0.99      0.97        71
     Maligno       0.98      0.93      0.95        43

    accuracy                           0.96       114
   macro avg       0.97      0.96      0.96       114
weighted avg       0.97      0.96      0.96       114

5. Modelamiento y Evaluacion (Over-Sampling)

Repetimos el proceso de entrenamiento y evaluacion utilizando el conjunto sobre-muestreado. Esto nos permite comparar directamente el efecto de cada estrategia de balanceo sobre el rendimiento de los modelos.

# Se obtiene X e y a partir del nuevo dataset
df_over_sampling = df_over_sampling.sample(frac = 1)
df_over_sampling_2 = df_over_sampling.iloc[:50000]

X_train_oversampling = df_over_sampling_2.drop('diagnosis', axis=1)
y_train_oversampling = df_over_sampling_2['diagnosis']

classifiers = {
    "Support Vector Classifier": SVC(),
    "DecisionTreeClassifier": DecisionTreeClassifier(),
    "Naive_bayes": GaussianNB()
}

for key, classifier in classifiers.items():
    classifier.fit(X_train_oversampling, y_train_oversampling)
    training_score = cross_val_score(classifier, X_train_oversampling, y_train_oversampling, cv=5)
    print("Clasificador: ", classifier.__class__.__name__, "tiene una precisión del", round(training_score.mean(), 2) * 100, "%")

Clasificador:  SVC tiene una precisión del 98.0 %
Clasificador:  DecisionTreeClassifier tiene una precisión del 94.0 %
Clasificador:  GaussianNB tiene una precisión del 93.0 %

#GridSearchCV para encontrar los mejores hiperparámetros

# Support Vector Classifier
svc_params = {'C': [0.5, 0.7, 0.9, 1], 'kernel': ['rbf', 'poly', 'sigmoid', 'linear']}
grid_svc = GridSearchCV(SVC(), svc_params)
grid_svc.fit(X_train_oversampling, y_train_oversampling)

# mejor modelo SVC  
svc = grid_svc.best_estimator_

# DecisionTree 
tree_params = {"criterion": ["gini", "entropy"], "max_depth": list(range(2,4,1)), 
              "min_samples_leaf": list(range(5,7,1))}
grid_tree = GridSearchCV(DecisionTreeClassifier(), tree_params)
grid_tree.fit(X_train_oversampling, y_train_oversampling)

# mejor modelo árbol de decisión  
tree_clf = grid_tree.best_estimator_

# Naive 
naive_clf = GaussianNB()
naive_clf.fit(X_train_oversampling, y_train_oversampling)

GaussianNB()

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

svc

SVC(C=0.7, kernel='linear')

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

tree_clf

DecisionTreeClassifier(max_depth=3, min_samples_leaf=6)

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

svc_score = cross_val_score(svc, X_train_oversampling, y_train_oversampling, cv=5)
print('Support Vector Classifier Cross Validation Score', round(svc_score.mean() * 100, 2).astype(str) + '%')

tree_score = cross_val_score(tree_clf, X_train_oversampling, y_train_oversampling, cv=5)
print('DecisionTree Classifier Cross Validation Score', round(tree_score.mean() * 100, 2).astype(str) + '%')

naive_score = cross_val_score(naive_clf, X_train_oversampling, y_train_oversampling, cv=5)
print('Naive Bayes Classifier Cross Validation Score', round(naive_score.mean() * 100, 2).astype(str) + '%')

Support Vector Classifier Cross Validation Score 97.9%
DecisionTree Classifier Cross Validation Score 92.84%
Naive Bayes Classifier Cross Validation Score 92.83%

from sklearn.metrics import roc_curve
from sklearn.model_selection import cross_val_predict
# Crea un DataFrame con todos los puntajes y los nombres de los clasificadores.

svc_pred = cross_val_predict(svc, X_train_oversampling, y_train_oversampling, cv=5,
                             method="decision_function")

tree_pred = cross_val_predict(tree_clf, X_train_oversampling, y_train_oversampling, cv=5)

naive_pred = cross_val_predict(naive_clf, X_train_oversampling, y_train_oversampling, cv=5)

from sklearn.metrics import roc_auc_score

print('Support Vector Classifier: ', roc_auc_score(y_train_oversampling, svc_pred))
print('Decision Tree Classifier: ', roc_auc_score(y_train_oversampling, tree_pred))
print('Naive Bayes Tree Classifier: ', roc_auc_score(y_train_oversampling, naive_pred))

Support Vector Classifier:  0.9936427209154481
Decision Tree Classifier:  0.9318181818181819
Naive Bayes Tree Classifier:  0.9283216783216784

6. Comparacion Final

Evaluamos los modelos entrenados con over-sampling sobre el conjunto de prueba y comparamos las matrices de confusion y metricas de clasificacion con los resultados obtenidos mediante under-sampling. Las curvas ROC nos permiten visualizar el trade-off entre sensibilidad y especificidad para cada modelo.

df_test = pd.DataFrame(Xtest, columns=dfbalanceado.columns[:-1])
df_test['diagnosis'] = ytest


print('Distribución de las clases en el conjunto de datos de submuestra - OVERSAMPLING')
print(df_test['diagnosis'].value_counts())



sns.countplot(x='diagnosis', data=df_test, palette=colors)
plt.title('Clases igualmente distribuidas', fontsize=14)
plt.show()

Distribución de las clases en el conjunto de datos de submuestra - OVERSAMPLING
diagnosis
0.0    71
1.0    43
Name: count, dtype: int64

from sklearn.metrics import confusion_matrix

y_pred_svc = svc.predict(Xtest)
y_pred_tree = tree_clf.predict(Xtest)
y_pred_naive = naive_clf.predict(Xtest)


svc_cf = confusion_matrix(ytest, y_pred_svc)
tree_cf = confusion_matrix(ytest, y_pred_tree)
naive_cf = confusion_matrix(ytest, y_pred_naive)

fig, ax = plt.subplots(2, 2,figsize=(22,12))


sns.heatmap(naive_cf, ax=ax[0][1], annot=True, cmap=plt.cm.copper)
ax[0][1].set_title("Naive Bayes \n Matriz de Confusión", fontsize=14)
ax[0][1].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[0][1].set_yticklabels(['', ''], fontsize=14, rotation=360)

sns.heatmap(svc_cf, ax=ax[1][0], annot=True, cmap=plt.cm.copper)
ax[1][0].set_title("Suppor Vector Machine \n Matriz de Confusión", fontsize=14)
ax[1][0].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[1][0].set_yticklabels(['', ''], fontsize=14, rotation=360)

sns.heatmap(tree_cf, ax=ax[1][1], annot=True, cmap=plt.cm.copper)
ax[1][1].set_title("DecisionTree \n Matriz de Confusión", fontsize=14)
ax[1][1].set_xticklabels(['', ''], fontsize=14, rotation=90)
ax[1][1].set_yticklabels(['', ''], fontsize=14, rotation=360)


plt.show()

y_true = [0, 1, 2, 2, 2]
y_pred = [0, 0, 2, 2, 1]
target_names = ['Benigno', 'Maligno']
print("Reporte clasificación SVC")
print(classification_report(ytest, y_pred_svc, target_names=target_names))
print("")
print("Reporte clasificación Árbol")
print(classification_report(ytest, y_pred_tree, target_names=target_names))
print("")
print("Reporte clasificación Naive Bayes")
print(classification_report(ytest, y_pred_naive, target_names=target_names))

Reporte clasificación SVC
              precision    recall  f1-score   support

     Benigno       0.97      0.99      0.98        71
     Maligno       0.98      0.95      0.96        43

    accuracy                           0.97       114
   macro avg       0.97      0.97      0.97       114
weighted avg       0.97      0.97      0.97       114


Reporte clasificación Árbol
              precision    recall  f1-score   support

     Benigno       0.97      0.97      0.97        71
     Maligno       0.95      0.95      0.95        43

    accuracy                           0.96       114
   macro avg       0.96      0.96      0.96       114
weighted avg       0.96      0.96      0.96       114


Reporte clasificación Naive Bayes
              precision    recall  f1-score   support

     Benigno       0.96      0.99      0.97        71
     Maligno       0.98      0.93      0.95        43

    accuracy                           0.96       114
   macro avg       0.97      0.96      0.96       114
weighted avg       0.97      0.96      0.96       114

svm_fpr,svm_tpr, trhreshold = roc_curve(ytest, y_pred_svc)
plt.plot(svm_fpr,svm_tpr, linestyle='-')

[<matplotlib.lines.Line2D at 0x7f3bf11ee510>]

svm_fpr,svm_tpr, trhreshold = roc_curve(ytest,y_pred_tree )
plt.plot(svm_fpr,svm_tpr, linestyle='-')

[<matplotlib.lines.Line2D at 0x7f3bf0d5f790>]

svm_fpr,svm_tpr, trhreshold = roc_curve(ytest,y_pred_naive )
plt.plot(svm_fpr,svm_tpr, linestyle='-')

[<matplotlib.lines.Line2D at 0x7f3bf0a25490>]

from sklearn.metrics import RocCurveDisplay

fig, ax = plt.subplots(figsize=(8, 6))
for name, clf in [("SVC", svc), ("Decision Tree", tree_clf), ("Naive Bayes", naive_clf)]:
    RocCurveDisplay.from_estimator(clf, Xtest, ytest, ax=ax, name=name)
ax.set_title("ROC Curves - Comparación de Clasificadores")
plt.show()