新聞分類系統的實現
1 系統開發工具和平臺
本文選擇Python作爲主要開發語言,作爲一個簡潔而又強大的腳步語言,Python整合了大量的第三方數據分析,算法處理框架,爲開發帶來極大的便利。
系統完整開發工具如圖所示:
圖1 分類系統開發工具彙總
在數據庫方面,選擇Mongodb來存儲爬取到的新聞信息。Mongodb作爲一個非關係型數據庫,只需將爬取到的新聞信息轉化爲鍵值對的類型便能完成存儲。在服務器構建上使用sqlite3作爲存儲數據庫,用於展示網頁上的新聞排序信息,sqlite3是python原生自帶的數據庫,使用sqlite3不用配置數據庫環境,只用將服務器框架與其鏈接,便能在本地配置一個簡易的數據庫。
在對新聞文本數據預處理上,使用python的Pandas讀取Mongodb數據庫中存儲信息,Pandas獨特的DataFrame數據結構,會調用Numpy和Scipy封裝的數據處理方法,將對數據表的處理轉化成類似數學上線性代數的矩陣運算,只用告知程序行列屬性,便能直接運算。相較C與Java對行列的複雜循環,Python的數據分析框架具有較高的代碼可讀性。但這會加大一定的運算時長作爲代價。對處理好的數據,使用Matpoltlib來完成繪圖。
與之同理,在處理中文分詞,和擬合算法模型上,使用Python的第三方庫jieba來作爲分詞工具,通過調用Scikit-learn封裝的機器學習算法來完成分類器模塊設計,而深度學習算法通過調用Keras封裝的算法模型來完成(Keras的後端爲TensorFlow)。在特徵提取上Word2vec算法會通過調用Gensim來實現。
在最後的服務器實現上,使用Django搭建了一個新聞發佈網站,前端使用Ajax請求來完成前後端交互,接收到新聞信息後,會通過Scikit-learn讀取本地保存的分類器PKL文件,來完成算法匹配。
運行電腦配置如下:
·CPU:(英特爾)Intel(R) Core(TM) i7-7700 CPU @ 3.60GHz(3601 MHz)
·內存:8.00 GB ( 2400 MHz)
·顯卡:Intel(R) HD Graphics 630 (1024 MB) (本項目使用的是cpu加速)
2 爬蟲模塊功能實現
如圖所示,實現爬蟲模塊主要以下5個步驟:
圖1爬蟲模塊整體流程
Step1:使用python自帶的urllib庫,對新浪新聞發送http請求,得到API的內的數據,實則是一串json格式封裝的新聞數據,包含新聞標題,新聞發佈時間,新聞鏈接,新聞評論人數,新聞來源等信息。
Step2:使用python自帶的json解析庫,解析json數據,得到需要的新聞標題,發佈時間,評論人數,新聞鏈接等信息。
Step3:異步加載通過訪問之前解析出來的新聞鏈接,爬取相應的新聞內容。
Step4:將所有數據存入mongodb數據庫中。爲之後的機器學習建模提供數據集。
Step5:使用多進程庫multiprocessing開啓進程池,使用多線程庫hreading開啓多線程。循環翻頁抓取全部所需信息。
由於API爬蟲的高效性,爬蟲程序直接訪問網頁端存儲的新聞信息Json數據,再配合多線程與多進程技術,爬蟲模塊半小時可以實現上百萬的新聞數據爬取。將爬取信息從數據庫導出結果如圖2所示:
圖2爬蟲導出結果
預處理模塊功能實現
通過pandas將數據庫的新聞信息讀入內存,數據格式爲dataframe,做數據預處理工作。爬來的新聞數據中,部分新聞內容信息是缺失的,我們將其從102萬條新聞數據中去除。最後可以使用的有48萬條新聞數據。
對這48萬條數據進行分組,總共分爲15個類的新聞數據。對不同新聞類別的統計如表所示:
可以發現lable 14 和lable 15的只具有幾百條,lable8和lable 11一個也沒有,數據分佈也十分不均勻。綜合考慮下,最後選擇剩下的11個lable,每個lable隨機抽取兩千條新聞信息。
每個類別分別代表:
汽車 | 財經 | IT | 健康 | 體育 | 旅遊 | 教育 | 軍事 | 文化 | 娛樂 | 時尚 |
1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 |
數據清洗好後,對每條新聞內容的長度進行統計發現新聞句子的長度基本在0至100內分佈。於是我們選擇100作爲原始的新聞數據截取長度。
(x:句子長度,y:新聞數量)
捨去長度高於100的那部分內容,只用前100字作爲訓練樣本,讓數據分別更均勻,同時也減少了特徵的維度。接下來,我們使用jieba分詞工具來處理這些新聞信息。
在使用jieba前,使用正則表達式將原文本信息轉換爲只有中文組成的句子,去除標點符號或者分割符對詞頻統計的影響。
Jieba分詞後效果如下:
Lable | 分詞示例 |
6 | 幸福 時刻 全家福 旅遊 照片 年 月 日 陝西省 翠 花山 全家福 旅遊 照片 年 月 日 ..... |
1 | 佛山市 佛 陳 大橋 下 週三 起 實施 全封閉 維修 廣佛 都市網 訊 佛山 日報 已經 ...... |
5 | 圖文 長沙 站 預選賽 頒獎儀式 鯊魚 球杆 樑寶忠 月 日 英倫 汽車 喬氏 杯 決戰 亨德利 ...... |
11 | 銀曼 專注 生活 細節 探尋 美麗 真諦 愛美 的 你 是否 曾 希望 每天 都 與 大自然 保持 ....... |
.... | ........ |
6 | 社區 小時 熱貼 推薦 歡迎 來到 河南 心情 時時 舒展 歡迎 加入 河南 版友 羣 |
分詞後需要去除停用詞,如圖展示的是對未去除停用詞的詞頻統計(TOP10):
可以發現在沒有去停用詞操作前,文檔中出現大量‘的’,‘在’,‘是’這類詞,但是其對分類貢獻率極低。通過去除停用詞我們可以得到下面結果:
4.5 分類器模塊功能實現
1 CNN
深度學習的CNN算法項目中使用到了python的第三方庫keras,keras爲使用TensorFlow提供的一定的便捷化的接口。它一定程度上降低了我們學習的難度,項目中使用keras可以便捷的構架神經網絡,而不用耗費大量的時間去學習TensorFlow的編碼解碼及構建特徵的編程方法。
在配置參數上我們選擇
100作爲每條新聞的最大長度,
單詞的向量空間維度爲200,
20%的數據作爲測試集
以及16%的數據作爲驗證集。
a. 不使用word2vec算法訓練cnn模型步驟:
·使用Tokenizer對所有文本數據做特徵提取。將新聞的文本數據轉化成由單詞的索引對應的序列。
·按配置參數的比例分割訓練集,驗證集,測試集。
·通過embedding技術對新聞特徵序列降維,生成100*200的二維向量矩陣。
· 設置1 層卷積層與池化層減少向量的長度,,通過一層 Flatten 層將 2 維向量壓縮到 1 維,最後由兩層Dense將向量長度收縮到 11 上,對應新聞數據集的 11 個類類別。
實驗結果如下:
訓練集準確率:0.8652
準確率爲:0.81450513
耗時:96s
對11個新聞類別的分類通過簡單的搭建一個神經網絡達到81%,但是從測試集準確度88%來看,存在一定的過擬合現象。總體來說在訓練的時間偏長,效率較低。
b. 使用word2vec算法的CNN模型步驟:
· 使用 word2vec 模型替代embedding層的1312 萬個參數。替換後embedding矩陣爲65604 x 200。65604表示65604個單詞。
· 其餘步驟如上所示。
實驗結果:
訓練集準確度:0.8629
測試集準確度:0.8262257
耗時:77s
模型的shape與之前一樣,過擬合現象減輕,準確率由原來的81%提升到了85.3%,這說明具備語義推理能力的word2vec可以一定程度上提高模型的準確率和運算性能。
2 LSTM
LSTM是深度學習算法中相對比較適合文本分類的一個模型,這裏以同樣的方法通過keras搭建LSTM網絡。
a. 不使用word2vec算法訓練LSTM模型
使用LSTM構架神經網絡的步驟和參數與CNN的相同,這裏不再做詳細說明。
在訓練集的準確率:0.7899
在測試集的準確率:0.7539733
耗時:161.9S
由於訓練的數據量偏小,LSTM並沒有發揮出其在自然語言處理上的優勢,另外使用LSTM訓練模型的時間爲CNN的2倍,略爲低效。
b. 使用word2vec算法的LSTM模型:
在訓練集的準確率:0.8746
在測試集的準確率:0.821892816
耗時:162S
間接說明新聞文檔的數量對LSTM的影響,使用word2vec產生的大量參數提升了語料庫的容量。使得準確率有所提升。
3.樸素貝葉斯
貝葉斯算法我們使用python的機器學習庫scikit-learn來完成,傳統的機器學習模型,我們不用將詞向量模型構建成二維的矩陣,來以分析圖像的思維來訓練文本,處理後的原數據矩陣爲65604*1。
a. 使用TF-IDF算法建立貝葉斯模型:
使用TF-IDF來做特徵提取,這裏我們使用CountVectorizer來建立特徵語料庫,語料庫的數量爲文檔出現所有詞的集合。
將測試集與訓練集分佈與之擬合:
因爲提前分割了數據集,所以在使用TF-IDF做特徵提取的時候需要主要詞對應語料庫的位置。應當使用總得語料庫來擬合訓練集與測試。如果分開擬合將導致訓練集與測試集相同索引對應不同單詞,從而造成較大的誤差。最後使用樸素貝葉斯算法建模,用測試集來驗證貝葉斯模型的準確率。
簡單的貝葉斯模型,準確率卻略高於LSTM,而且在運行時間上,貝葉斯模型的運行時間不到2S,是LSTM的90分之一。
b. 使用word2vec訓練貝葉斯模型:
配置參數爲,100的新聞文章最大長度,N爲4的字流窗大小,使用多核cpu加速,使用skip-gram做特徵提取,迭代次數爲10次。
Word2vel處理後的文本矩陣:
每個字對應了大量的權重,且由於word2vec賦予產生的矩陣一種連續性,使用樸素貝葉斯將不能處理這些連續得向量矩陣,這裏我們使用高斯貝葉斯,假設矩陣滿足正態分佈。
訓練結果如下:
訓練正確率出其的低,且耗時爲樸素貝葉斯模型的22倍。從直觀的想法中,樸素貝葉斯正確率偏高,而使用word2vec將貝葉斯的單詞獨立性的假設得到了補充,按理應該正確率得到提升。下一章實驗結果展示將會討論這個問題。
4. Svm
Svm處理思路與樸素貝葉斯相同,這裏指明一下,項目中SVM使用線性核。這是對比過高斯與多項式得出的。
a. 使用TF-IDF算法建立SVM模型:
由於特徵構建原理相同,這裏不做解釋,詳細可以參考上面貝葉斯的文檔。
正確率:84.4% 雖不及貝葉斯,但是整體效果還是不錯的。
b. 使用word2vec算法建立SVM模型:
正確率只有77.08%,但是相較word2vec在高斯貝葉斯模型模型的表現已經很好了。相同參數下,運行時間上達到了134.9s,效率很低。
4.4系統界面的實現
系統界面使用python的Django服務器框架開發,使用python原生的sqlite3作爲數據庫支持。前端使用JavaScript,CSS建立了一個較爲簡約的UI界面。主頁建立了一個發佈按鈕。
點擊發布按鈕(1),產生如圖一個彈框:
彈框中,添加鏈接這欄可以添加一個URL(2),因爲本系統是針對新浪新聞而開發的,使用目前僅支持對新浪新聞相應鏈接的提取。添加鏈接後,點擊獲取標題(3),可以自動爬取對應的新聞標題和對應的摘要(4),點擊發布按鈕(5)。系統後臺接收新聞標題與對應的摘要,經過分詞後,通過與TF-IDF算法特徵提取保存的單詞及對應TF-IDF值矩陣匹配,轉化成向量矩陣。這個向量矩陣,再與我們離線訓練出來的貝葉斯模型,通過貝葉斯算法計算其所屬類別的概率,達到預測的目的。
觀察上圖可以發現,貝葉斯算法具有較高的準確率,可以做到良好的分類效果。
第五章 實驗結果分析
5.1 系統評估指標(ROC,AUC,訓練所需時間)
a. ROC,AUC
在介紹系統評估指標前,我們先了解4個概念:
·True Positive(TP):意思是對於某一個類別的新聞信息,算法對新聞信息做出預測,且預測類別與此類別相同,TP的值表示預測該類別相同的個數。
·False Positive(FP):數值表示預測某一類別預測類別與真實類別不同的個數;
·True Negative(TN):數值表示預測某一類別預測爲此類別,但是真實值非此類別的個數;
·False Negative(FN):數值表示預測某一類別預測爲非此類別,且真實值也非此類別的個數;
基於此,我們就可以計算出準確率(precision rate)、召回率(recall rate)。
|
| 預測類別 | |
真實結果 |
| 財經 | 非財經 |
財經 | 170(TP) | 300(FN) | |
非財經 | 30(FP) | 1700(TN) |
以上表爲例,TP的值爲200,FN的值爲30,FP的值爲300,TN的值爲2000。
那麼, 準確率=170/(170+30) = 85% ,召回率=170/(170+300)= 36.17%。
ROC曲線就是準確率隨召回率的變化情況。ROC曲線越接近左上角,分類效果越好。AUC曲線表示ROC曲線下的面積,AUC面積越大,分類效果越好
5.2 算法擬合數據集說明
爲了讓算法能夠均勻計算到每個類別,項目數據集上選擇從每個類別中隨機抽取2000條新聞數據進行算法擬合。其中文化類別只爬取到1924條,所以使用對文化類別選取全部的1924條作爲樣本。
類別字典映射表如下:
汽車 | 財經 | IT | 健康 | 體育 | 旅遊 | 教育 | 軍事 | 文化 | 娛樂 | 時尚 |
1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 |
在深度學習算法CNN,LSTM中選擇20%的數據作爲測試集以及16%的數據作爲驗證集。
在傳統機器學習算法中,我們選擇80%的數據作爲訓練集,20%的數據作爲驗證集。
5.3 分類系統算法評估
算法 | 準確度 | 訓練所需時間 |
樸素貝葉斯+TF-IDF | 0.85430157 | <2s |
SVM+TF-IDF | 0.84435707 | 45s |
CNN+word2vec | 0.8262257 | 77s |
LSTM+word2vec | 0.821892816 | 172s |
CNN+Tokenizer | 0.8150513 | 96s |
SVM+word2vec | 0.77081406 | 23s |
LSTM+Tokenizer | 0.7539733 | 161.9s |
樸素貝葉斯+word2vec | 0.5790934 | 13s |
樸素貝葉斯算法+TF-IDF ROC曲線:
SVM算法+TF-IDF ROC曲線:
CNN算法word2vec ROC曲線:
LSTM算法word2vec ROC曲線:
通過對比訓練所需時間,樸素貝葉斯算法配合TF-IDF只需要2S就能完成對1萬4千多條信息的數學建模。且對比不同算法模型的ROC圖,貝葉斯算法綜合下來ROC曲線最接近左上角,分類效果最好。ROC曲線下的面積對比中,樸素貝葉斯AUC面積最大大,分類效果最好。
由此判斷,樸素貝葉斯配合TF-IDF是最爲適合作爲本項目新聞分類器的算法模型。其算法成功率爲0.85430157,也支持對大量新聞數據的分類預測。
源碼:
深度學習使用到的語料庫爲維基百科訓練出來的語料庫。
CNN:
#coding:utf-8
import sys
import keras
import matplotlib.pyplot as plt
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 200
VALIDATION_SPLIT = 0.16
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_texts = train_texts + test_texts
all_labels = train_labels + test_labels
print ('(2) doc to var...')
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
import numpy as np
tokenizer = Tokenizer()
tokenizer.fit_on_texts(all_texts)
sequences = tokenizer.texts_to_sequences(all_texts)
word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))
data = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)
labels = to_categorical(np.asarray(all_labels))
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
print ('(3) split data set...')
# split the data into training set, validation set, and test set
p1 = int(len(data)*(1-VALIDATION_SPLIT-TEST_SPLIT))
p2 = int(len(data)*(1-TEST_SPLIT))
x_train = data[:p1]
y_train = labels[:p1]
x_val = data[p1:p2]
y_val = labels[p1:p2]
x_test = data[p2:]
y_test = labels[p2:]
print ('train docs: '+str(len(x_train)))
print ('val docs: '+str(len(x_val)))
print ('test docs: '+str(len(x_test)))
print ('(5) training model...')
from keras.layers import Dense, Input, Flatten, Dropout
from keras.layers import Conv1D, MaxPooling1D, Embedding, GlobalMaxPooling1D
from keras.models import Sequential
model = Sequential()
model.add(Embedding(len(word_index) + 1, EMBEDDING_DIM, input_length=MAX_SEQUENCE_LENGTH))
model.add(Dropout(0.2))
model.add(Conv1D(250, 3, padding='valid', activation='relu', strides=1))
model.add(MaxPooling1D(3))
model.add(Flatten())
model.add(Dense(EMBEDDING_DIM, activation='relu'))
model.add(Dense(labels.shape[1], activation='softmax'))
model.summary()
#plot_model(model, to_file='model.png',show_shapes=True)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['acc'])
print (model.metrics_names)
model.fit(x_train, y_train, callbacks=[history],validation_data=(x_val, y_val), epochs=2, batch_size=128)
#model.save('cnn.h5')
print ('(6) testing model...')
print (model.evaluate(x_test, y_test))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = model.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
CNN+word2vec
#coding:utf-8
import sys
import keras
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 128
VALIDATION_SPLIT = 0.16
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_texts = train_texts + test_texts
all_labels = train_labels + test_labels
print ('(2) doc to var...')
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
import numpy as np
tokenizer = Tokenizer()
tokenizer.fit_on_texts(all_texts)
sequences = tokenizer.texts_to_sequences(all_texts)
word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))
data = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)
labels = to_categorical(np.asarray(all_labels))
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
print ('(3) split data set...')
# split the data into training set, validation set, and test set
p1 = int(len(data)*(1-VALIDATION_SPLIT-TEST_SPLIT))
p2 = int(len(data)*(1-TEST_SPLIT))
x_train = data[:p1]
y_train = labels[:p1]
x_val = data[p1:p2]
y_val = labels[p1:p2]
x_test = data[p2:]
y_test = labels[p2:]
print ('train docs: '+str(len(x_train)))
print ('val docs: '+str(len(x_val)))
print ('test docs: '+str(len(x_test)))
print ('(4) load word2vec as embedding...')
import gensim
from keras.utils import plot_model
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(VECTOR_DIR, binary=True)
embedding_matrix = np.zeros((len(word_index) + 1, EMBEDDING_DIM))
not_in_model = 0
in_model = 0
for word, i in word_index.items():
if word in w2v_model:
in_model += 1
embedding_matrix[i] = np.asarray(w2v_model[word], dtype='float32')
else:
not_in_model += 1
print (str(not_in_model)+' words not in w2v model')
from keras.layers import Embedding
embedding_layer = Embedding(len(word_index) + 1,
EMBEDDING_DIM,
weights=[embedding_matrix],
input_length=MAX_SEQUENCE_LENGTH,
trainable=False)
print ('(5) training model...')
from keras.layers import Dense, Input, Flatten, Dropout
from keras.layers import Conv1D, MaxPooling1D, Embedding, GlobalMaxPooling1D
from keras.models import Sequential
model = Sequential()
model.add(embedding_layer)
model.add(Dropout(0.2))
model.add(Conv1D(250, 3, padding='valid', activation='relu', strides=1))
model.add(MaxPooling1D(3))
model.add(Flatten())
model.add(Dense(EMBEDDING_DIM, activation='relu'))
model.add(Dense(labels.shape[1], activation='softmax'))
model.summary()
#plot_model(model, to_file='model.png',show_shapes=True)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['acc'])
print( model.metrics_names)
model.fit(x_train, y_train, validation_data=(x_test, y_test), epochs=2, batch_size=128)
model.save('word_vector_cnn.h5')
print ('(6) testing model...')
print (model.evaluate(x_test, y_test))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = model.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
LSTM
#coding:utf-8
import keras
import matplotlib.pyplot as plt
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 200
VALIDATION_SPLIT = 0.16
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_texts = train_texts + test_texts
all_labels = train_labels + test_labels
print ('(2) doc to var...')
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
import numpy as np
tokenizer = Tokenizer()
tokenizer.fit_on_texts(all_texts)
sequences = tokenizer.texts_to_sequences(all_texts)
word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))
data = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)
labels = to_categorical(np.asarray(all_labels))
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
print ('(3) split data set...')
p1 = int(len(data)*(1-VALIDATION_SPLIT-TEST_SPLIT))
p2 = int(len(data)*(1-TEST_SPLIT))
x_train = data[:p1]
y_train = labels[:p1]
x_val = data[p1:p2]
y_val = labels[p1:p2]
x_test = data[p2:]
y_test = labels[p2:]
print ('train docs: '+str(len(x_train)))
print ('val docs: '+str(len(x_val)))
print ('test docs: '+str(len(x_test)))
from keras.layers import Dense, Input, Flatten, Dropout
from keras.layers import LSTM, Embedding
from keras.models import Sequential
model = Sequential()
model.add(Embedding(len(word_index) + 1, EMBEDDING_DIM, input_length=MAX_SEQUENCE_LENGTH))
model.add(LSTM(200, dropout=0.2, recurrent_dropout=0.2))
model.add(Dropout(0.2))
model.add(Dense(labels.shape[1], activation='softmax'))
model.summary()
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['acc'])
history = model.fit(x_train, y_train,validation_data=(x_val, y_val), epochs=2, batch_size=128)
#model.save('lstm.h5')
print (model.evaluate(x_test, y_test))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = model.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
LSTM+word2vec
#coding:utf-8
import sys
import keras
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 128
VALIDATION_SPLIT = 0.16
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_texts = train_texts + test_texts
all_labels = train_labels + test_labels
print ('(2) doc to var...')
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
import numpy as np
tokenizer = Tokenizer()
tokenizer.fit_on_texts(all_texts)
sequences = tokenizer.texts_to_sequences(all_texts)
word_index = tokenizer.word_index
print('Found %s unique tokens.' % len(word_index))
data = pad_sequences(sequences, maxlen=MAX_SEQUENCE_LENGTH)
labels = to_categorical(np.asarray(all_labels))
print('Shape of data tensor:', data.shape)
print('Shape of label tensor:', labels.shape)
print ('(3) split data set...')
p1 = int(len(data)*(1-VALIDATION_SPLIT-TEST_SPLIT))
p2 = int(len(data)*(1-TEST_SPLIT))
x_train = data[:p1]
y_train = labels[:p1]
x_val = data[p1:p2]
y_val = labels[p1:p2]
x_test = data[p2:]
y_test = labels[p2:]
print ('train docs: '+str(len(x_train)))
print ('val docs: '+str(len(x_val)))
print ('test docs: '+str(len(x_test)))
print ('(4) load word2vec as embedding...')
import gensim
from keras.utils import plot_model
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(VECTOR_DIR, binary=True)
embedding_matrix = np.zeros((len(word_index) + 1, EMBEDDING_DIM))
not_in_model = 0
in_model = 0
for word, i in word_index.items():
if word in w2v_model:
in_model += 1
embedding_matrix[i] = np.asarray(w2v_model[word], dtype='float32')
else:
not_in_model += 1
print (str(not_in_model)+' words not in w2v model')
from keras.layers import Embedding
embedding_layer = Embedding(len(word_index) + 1,
EMBEDDING_DIM,
weights=[embedding_matrix],
input_length=MAX_SEQUENCE_LENGTH,
trainable=False)
print ('(5) training model...')
from keras.layers import Dense, Input, Flatten, Dropout
from keras.layers import LSTM, Embedding
from keras.models import Sequential
model = Sequential()
model.add(embedding_layer)
model.add(LSTM(200, dropout=0.2, recurrent_dropout=0.2))
model.add(Dropout(0.2))
model.add(Dense(labels.shape[1], activation='softmax'))
model.summary()
#plot_model(model, to_file='model.png',show_shapes=True)
model.compile(loss='categorical_crossentropy',
optimizer='rmsprop',
metrics=['acc'])
print (model.metrics_names)
model.fit(x_train, y_train, validation_data=(x_val, y_val), epochs=2, batch_size=128)
#model.save('word_vector_lstm.h5')
print ('(6) testing model...')
print (model.evaluate(x_test, y_test))
#畫圖
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = model.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
樸素貝葉斯
#coding:utf-8
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.naive_bayes import MultinomialNB
from sklearn import metrics
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_text = train_texts + test_texts
count_v0= CountVectorizer();
counts_all = count_v0.fit_transform(all_text);
count_v1= CountVectorizer(vocabulary=count_v0.vocabulary_);
counts_train = count_v1.fit_transform(train_texts);
print ("the shape of train is "+repr(counts_train.shape) )
count_v2 = CountVectorizer(vocabulary=count_v0.vocabulary_);
counts_test = count_v2.fit_transform(test_texts);
print ("the shape of test is "+repr(counts_test.shape) )
tfidftransformer = TfidfTransformer();
train_data = tfidftransformer.fit(counts_train).transform(counts_train);
test_data = tfidftransformer.fit(counts_test).transform(counts_test);
x_train = train_data
y_train = train_labels
x_test = test_data
y_test = test_labels
clf = MultinomialNB(alpha = 0.01)
clf.fit(x_train, y_train);
preds = clf.predict(x_test);
num = 0
preds = preds.tolist()
for i,pred in enumerate(preds):
if int(pred) == int(y_test[i]):
num += 1
print ('precision_score:' + str(float(num) / len(preds)))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = clf.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
高斯貝葉斯+word2vec
#coding:utf-8
from sklearn.naive_bayes import MultinomialNB
from sklearn.preprocessing import scale
from sklearn.naive_bayes import GaussianNB
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 128
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_docs = open('train_contents.txt',encoding = 'utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding = 'utf-8').read().split('\n')
test_docs = open('test_contents.txt',encoding = 'utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding = 'utf-8').read().split('\n')
print ('(2) doc to var...')
import gensim
import numpy as np
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(VECTOR_DIR, binary=True)
def buildWordVector(text, size):
'''
利用函數獲得每個文本中所有詞向量的平均值來表徵該特徵向量。
'''
vec = np.zeros(128).reshape((1, size))
count = 0
for word in text:
try:
vec += w2v_model[word].reshape((1, 128))
count += 1
except KeyError:
continue
if count != 0:
vec /= count
return vec
'''獲取需要所有文檔的詞向量,並且標準化出來'''
x_train1 = np.concatenate([buildWordVector(x, 128) for x in train_docs])
print ("the shape of train is "+repr(x_train1.shape) )
x_train = scale(x_train1)
x_test1 = np.concatenate([buildWordVector(x, 128) for x in test_docs])
print ("the shape of train is "+repr(x_test1.shape) )
x_test = scale(x_test1)
y_train = train_labels
y_test = test_labels
clf = GaussianNB()
clf.fit(x_train, y_train);
preds = clf.predict(x_test);
num = 0
preds = preds.tolist()
for i,pred in enumerate(preds):
if int(pred) == int(y_test[i]):
num += 1
print ('precision_score:' + str(float(num) / len(preds)))
SVM
#coding:utf-8
import sys
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 200
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_texts = open('train_contents.txt',encoding='utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding='utf-8').read().split('\n')
test_texts = open('test_contents.txt',encoding='utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding='utf-8').read().split('\n')
all_text = train_texts + test_texts
print ('(2) doc to var...')
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
count_v0= CountVectorizer();
counts_all = count_v0.fit_transform(all_text);
count_v1= CountVectorizer(vocabulary=count_v0.vocabulary_);
counts_train = count_v1.fit_transform(train_texts);
print ("the shape of train is "+repr(counts_train.shape) )
count_v2 = CountVectorizer(vocabulary=count_v0.vocabulary_);
counts_test = count_v2.fit_transform(test_texts);
print ("the shape of test is "+repr(counts_test.shape) )
tfidftransformer = TfidfTransformer();
train_data = tfidftransformer.fit(counts_train).transform(counts_train);
test_data = tfidftransformer.fit(counts_test).transform(counts_test);
x_train = train_data
y_train = train_labels
x_test = test_data
y_test = test_labels
print ('(3) SVM...')
from sklearn.svm import SVC
svclf = SVC(kernel = 'linear')
svclf.fit(x_train,y_train)
preds = svclf.predict(x_test);
num = 0
preds = preds.tolist()
for i,pred in enumerate(preds):
if int(pred) == int(y_test[i]):
num += 1
print ('precision_score:' + str(float(num) / len(preds)))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = clf.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()
SVM+word2vec
#coding:utf-8
from sklearn.preprocessing import scale
VECTOR_DIR = 'vectors.bin'
MAX_SEQUENCE_LENGTH = 100
EMBEDDING_DIM = 200
TEST_SPLIT = 0.2
print ('(1) load texts...')
train_docs = open('train_contents.txt',encoding = 'utf-8').read().split('\n')
train_labels = open('train_labels.txt',encoding = 'utf-8').read().split('\n')
test_docs = open('test_contents.txt',encoding = 'utf-8').read().split('\n')
test_labels = open('test_labels.txt',encoding = 'utf-8').read().split('\n')
print ('(2) doc to var...')
import gensim
import numpy as np
w2v_model = gensim.models.KeyedVectors.load_word2vec_format(VECTOR_DIR, binary=True)
def buildWordVector(text, size):
'''
利用函數獲得每個文本中所有詞向量的平均值來表徵該特徵向量。
'''
vec = np.zeros(128).reshape((1, size))
count = 0
for word in text:
try:
vec += w2v_model[word].reshape((1, 128))
count += 1
except KeyError:
continue
if count != 0:
vec /= count
return vec
'''獲取需要所有文檔的詞向量,並且標準化出來'''
x_train1 = np.concatenate([buildWordVector(x, 128) for x in train_docs])
x_train = scale(x_train1)
x_test1 = np.concatenate([buildWordVector(x, 128) for x in test_docs])
x_test = scale(x_test1)
y_train = train_labels
y_test = test_labels
print ('(3) SVM...')
from sklearn.svm import SVC
svclf = SVC(kernel = 'linear')
svclf.fit(x_train,y_train)
preds = svclf.predict(x_test);
num = 0
preds = preds.tolist()
for i,pred in enumerate(preds):
if int(pred) == int(y_test[i]):
num += 1
print ('precision_score:' + str(float(num) / len(preds)))
import matplotlib.pyplot as plt
from sklearn.metrics import roc_curve,auc
import numpy as np
from scipy import interp
y_score = clf.predict(x_test)
lw = 2
n_classes = 11
fpr = dict()
tpr = dict()
roc_auc = dict()
for i in range(n_classes):
fpr[i], tpr[i], _ = roc_curve(y_test[:, i], y_score[:, i])
roc_auc[i] = auc(fpr[i], tpr[i])
# Compute micro-average ROC curve and ROC area
fpr["micro"], tpr["micro"], _ = roc_curve(y_test.ravel(), y_score.ravel())
roc_auc["micro"] = auc(fpr["micro"], tpr["micro"])
# Compute macro-average ROC curve and ROC area
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(n_classes)]))
# Then interpolate all ROC curves at this points
mean_tpr = np.zeros_like(all_fpr)
for i in range(n_classes):
mean_tpr += interp(all_fpr, fpr[i], tpr[i])
# Finally average it and compute AUC
mean_tpr /= n_classes
fpr["macro"] = all_fpr
tpr["macro"] = mean_tpr
roc_auc["macro"] = auc(fpr["macro"], tpr["macro"])
# Plot all ROC curves
plt.figure()
plt.plot(fpr["micro"], tpr["micro"],
label='micro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["micro"]),
color='deeppink', linestyle=':', linewidth=4)
plt.plot(fpr["macro"], tpr["macro"],
label='macro-average ROC curve (area = {0:0.2f})'
''.format(roc_auc["macro"]),
color='navy', linestyle=':', linewidth=4)
colors = cycle(['aqua', 'darkorange', 'cornflowerblue'])
for i, color in zip(range(n_classes), colors):
plt.plot(fpr[i], tpr[i], color=color, lw=lw,
label='ROC curve of class {0} (area = {1:0.2f})'
''.format(i, roc_auc[i]))
plt.plot([0, 1], [0, 1], 'k--', lw=lw)
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Some extension of Receiver operating characteristic to multi-class')
plt.legend(loc="lower right")
plt.show()