👩‍💻

言語処理100本ノック 2020 (Rev 2) 第6章: 機械学習 56. 適合率,再現率,F1スコアの計測

2023/01/14に公開約4,700字

問題

56. 適合率,再現率,F1スコアの計測

52で学習したロジスティック回帰モデルの適合率,再現率,F1スコアを,評価データ上で計測せよ.カテゴリごとに適合率,再現率,F1スコアを求め,カテゴリごとの性能をマイクロ平均(micro-average)とマクロ平均(macro-average)で統合せよ.

solution56.py
import pandas as pd
import numpy as np
import re
import seaborn as sns
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
from sklearn.metrics import confusion_matrix
from sklearn.metrics import precision_score, recall_score, f1_score

def preprocessing(data):
    x = []
    y = []
    label = {"b":0, "e":1, "t":2, "m":3}

    for title, category in data:
        title = re.sub("[0-9]+", "0", title)
        x.append(title.lower())
        y.append(label[category])

    return x,y

def score(model, X):
    pred = model.predict([X])
    pred_proba = model.predict_proba([X])[0, pred]
    return pred[0], pred_proba[0]

def calc(Y, pred):
    ppv = precision_score(Y, pred, average=None)
    ppv_micro = precision_score(Y, pred, average="micro").reshape(1)
    ppv_macro = precision_score(Y, pred, average="macro").reshape(1)
    ppv = np.concatenate([ppv, ppv_micro, ppv_macro])

    recall = recall_score(Y, pred, average=None)
    recall_micro = recall_score(Y, pred, average="micro").reshape(1)
    recall_macro = recall_score(Y, pred, average="macro").reshape(1)
    recall = np.concatenate([recall, recall_micro, recall_macro])

    f1 = f1_score(Y, pred, average=None)
    f1_micro = f1_score(Y, pred, average="micro").reshape(1)
    f1_macro = f1_score(Y, pred, average="macro").reshape(1)
    f1 = np.concatenate([f1, f1_micro, f1_macro])


    index = ["0", "1", "2", "3", "micro", "macro"]
    scores = pd.DataFrame({"ppv": ppv, "recall": recall, "f1":f1}, index=index)

    return scores

np.random.seed(123)
df = pd.read_csv("chapter06/NewsAggregatorDataset/newsCorpora.csv", header=None, sep="\t", names=["ID", "TITLE", "URL", "PUBLISHER", "CATEGORY", "STORY", "HOSTNAME", "TIMESTAMP"])
df = df.loc[df["PUBLISHER"].isin(["Reuters", "Huffington Post", "Businessweek", "Contactmusic.com", "Daily Mail"]), ["TITLE", "CATEGORY"]]

train, valid_test = train_test_split(df, test_size=0.2, shuffle=True)
valid, test = train_test_split(valid_test, test_size=0.5, shuffle=True)

train = np.array(train)
valid = np.array(valid)
test = np.array(test)

X_train, Y_train = preprocessing(train)
X_valid, Y_valid = preprocessing(valid)
X_test, Y_test = preprocessing(test)

tfidfvectorizer = TfidfVectorizer(min_df=0.001)

X_train = tfidfvectorizer.fit_transform(X_train).toarray()
X_valid = tfidfvectorizer.transform(X_valid).toarray()
X_test = tfidfvectorizer.transform(X_test).toarray()

model = LogisticRegression()
model.fit(X_train, Y_train)

train_pred = []
test_pred = []

for X in X_train:
    train_pred.append(score(model, X)[0])

for X in X_test:
    test_pred.append(score(model, X)[0])

print(calc(Y_test, test_pred))
output
            ppv    recall        f1
0      0.885572  0.930314  0.907392
1      0.861060  0.949597  0.903164
2      0.808333  0.584337  0.678322
3      0.859375  0.561224  0.679012
micro  0.867316  0.867316  0.867316
macro  0.853585  0.756368  0.791972

calc関数では、与えられた正解ラベルYと予測ラベルpredに対して、PrecisionRecallF1スコアを計算し、それらをまとめたデータフレームを返すことです。

ppv(Precision)を計算するために、precision_score(Y, pred, average=None)を使用し、各クラスごとのPrecisionを計算します。ppv(Precision)のマイクロ平均を計算するために、precision_score(Y, pred, average="micro").reshape(1)を使用し、マイクロ平均を計算します。ppv(Precision)のマクロ平均を計算するために、precision_score(Y, pred, average="macro").reshape(1)を使用し、マクロ平均を計算します。

上記で計算した各クラスごとのPrecisionマイクロ平均マクロ平均を結合し、ppv配列を作成します。以上の処理をrecall, f1分も行います。

https://github.com/kurokawa5/nlp100_2020/blob/main/chapter06/solution56.py

参考記事

第6章: 機械学習
pandas.DataFrameの行を条件で抽出するquery
scikit-learnでデータを訓練用とテスト用に分割するtrain_test_split
Numpy で乱数を生成する
sklearn.feature_extraction.text.TfidfVectorizer
sklearn.linear_model.LogisticRegression
scikit-learnで混同行列を生成、適合率・再現率・F1値などを算出
NumPy配列ndarrayを結合(concatenate, stack, blockなど)

Discussion

ログインするとコメントできます