다항분류

실습 준비

데이터

from sklearn.datasets import fetch_20newsgroups
news = fetch_20newsgroups()
news['target_names']

['alt.atheism',
 'comp.graphics',
 'comp.os.ms-windows.misc',
 'comp.sys.ibm.pc.hardware',
 'comp.sys.mac.hardware',
 'comp.windows.x',
 'misc.forsale',
 'rec.autos',
 'rec.motorcycles',
 'rec.sport.baseball',
 'rec.sport.hockey',
 'sci.crypt',
 'sci.electronics',
 'sci.med',
 'sci.space',
 'soc.religion.christian',
 'talk.politics.guns',
 'talk.politics.mideast',
 'talk.politics.misc',
 'talk.religion.misc']

문서 단어 행렬

from sklearn.feature_extraction.text import CountVectorizer
cv = CountVectorizer(min_df=0.01, max_df=0.5, stop_words='english')
dtm = cv.fit_transform(news['data'])

dtm.shape

(11314, 1916)

x와 y 지정

x = dtm
y = news['target']

분할

from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(
    x, y, test_size=0.2, random_state=42)

Logistic Regression

사이킷런의 로지스틱 회귀분석은 OvR을 지원

from sklearn.linear_model import LogisticRegression

model = LogisticRegression(multi_class='ovr', max_iter=200)
model.fit(x_train, y_train)

LogisticRegression(max_iter=200, multi_class='ovr')

LogisticRegression(max_iter=200, multi_class='ovr')

multi_class=’ovr’로 설정(현재 기본값이 OvR이므로 생략해도 무방)

prob = model.predict_proba(x_test)
prob[0]

array([2.06979569e-01, 4.98778872e-04, 1.51227474e-01, 2.37753022e-01,
       5.87049790e-02, 5.38452734e-03, 2.00523865e-03, 2.88793490e-03,
       6.22238404e-03, 2.05132109e-03, 3.60580256e-03, 4.32783621e-04,
       5.88802289e-02, 1.48888125e-03, 1.95599123e-03, 9.33936922e-05,
       3.85273142e-02, 2.82426090e-04, 1.56440433e-03, 2.19453545e-01])

확률이 가장 높은 클래스 보기

import numpy as np
np.argmax(prob, axis=1)

array([ 3,  2,  6, ..., 14, 16, 11], dtype=int64)

확률 최대값 보기

np.max(prob, axis=1)

array([0.23775302, 0.95134806, 0.86915966, ..., 0.91493761, 0.83797877,
       0.91793744])

정확도 평가

model.score(x_test, y_test)

0.8020326999558108

가중치 표 만들기

import pandas as pd
wd = pd.DataFrame(model.coef_.T, 
                  index=cv.get_feature_names_out(),
                  columns=news['target_names'])

자동차 게시판의 가중치가 가장 높은 단어들

wd['rec.autos'].sort_values().tail(10)

models     1.103112
early      1.103447
89         1.117528
engine     1.170994
owners     1.263317
dealer     1.273244
auto       1.439482
warning    1.603697
cars       1.895346
car        2.098276
Name: rec.autos, dtype: float64

나이브 베이즈

from sklearn.naive_bayes import BernoulliNB
model = BernoulliNB()
model.fit(x_train, y_train)
model.score(x_test, y_test)

0.7061422889969068

텐서플로

import tensorflow as tf

tf.nn.softmax([-1. ,  0.5,  2. ])

<tf.Tensor: shape=(3,), dtype=float32, numpy=array([0.03911257, 0.17529039, 0.785597  ], dtype=float32)>

x = 2.0
tf.nn.sigmoid(x)

<tf.Tensor: shape=(), dtype=float32, numpy=0.8807971>

tf.nn.softmax([0, x])

<tf.Tensor: shape=(2,), dtype=float32, numpy=array([0.11920291, 0.880797  ], dtype=float32)>

from tensorflow.keras.layers import Dense

model = tf.keras.Sequential([
    Dense(20, activation='softmax')
])

model.compile(loss='sparse_categorical_crossentropy', metrics=['accuracy'])
model.fit(x_train.A, y_train, epochs=10)

Epoch 1/10
283/283 [==============================] - 0s 887us/step - loss: 2.0657 - accuracy: 0.5386
Epoch 2/10
283/283 [==============================] - 0s 910us/step - loss: 1.1081 - accuracy: 0.8181
Epoch 3/10
283/283 [==============================] - 0s 933us/step - loss: 0.7881 - accuracy: 0.8624
Epoch 4/10
283/283 [==============================] - 0s 899us/step - loss: 0.6315 - accuracy: 0.8847
Epoch 5/10
283/283 [==============================] - 0s 903us/step - loss: 0.5391 - accuracy: 0.8971
Epoch 6/10
283/283 [==============================] - 0s 915us/step - loss: 0.4741 - accuracy: 0.9055
Epoch 7/10
283/283 [==============================] - 0s 906us/step - loss: 0.4241 - accuracy: 0.9157
Epoch 8/10
283/283 [==============================] - 0s 919us/step - loss: 0.3871 - accuracy: 0.9213
Epoch 9/10
283/283 [==============================] - 0s 900us/step - loss: 0.3579 - accuracy: 0.9266
Epoch 10/10
283/283 [==============================] - 0s 981us/step - loss: 0.3330 - accuracy: 0.9317

<keras.callbacks.History at 0x2e6fd331c60>

model.evaluate(x_test.A, y_test)

71/71 [==============================] - 0s 879us/step - loss: 0.6993 - accuracy: 0.8188

[0.6993224620819092, 0.8188245892524719]