我正在研究一个多类文本分类问题,它必须提供前5个匹配项,而不是最好的匹配项。因此,成功被定义为前5个匹配中至少有一个是正确的分类。该算法必须达到至少95%的成功率给我们如何定义成功以上。当然,我们将在数据的子集上训练我们的模型,并在剩余的子集上进行测试,以验证我们的模型的成功。
我一直在使用pythons scikit learns predict_proba()函数来选择前5个匹配项,并使用自定义脚本计算下面的成功率,该脚本似乎在我的示例数据上运行得很好,但是,我注意到前5个成功率低于在我自己的自定义数据上使用.predict()的前1个成功率,后者在数学上是不可能的。这是因为排名第一的结果将自动包含在排名前5的结果中,因此成功率必须至少等于排名前1的成功率(如果不是更多的话)。为了解决问题,我使用predict()与predict_proba()比较前1名的成功率,以确保它们相等,并确保前5名的成功率大于前1名。
我已经设置了下面的脚本来引导您浏览我的逻辑,看看是否在某个地方做出了错误的假设,或者我的数据是否有问题需要修复。我正在测试许多分类器和特征,但为了简单起见,您将看到我只是使用计数向量作为特征,使用logistic回归作为分类器,因为我不相信(据我所知,这是问题的一部分)。
如果有人能解释我为什么会发现这种差异,我将不胜感激。
代码:
# Set up environment
from sklearn.datasets import fetch_20newsgroups
from sklearn.linear_model import LogisticRegression
from sklearn import metrics, model_selection
from sklearn.feature_extraction.text import CountVectorizer
import pandas as pd
import numpy as np
#Read in data and do just a bit of preprocessing
# User's Location of git repository
Git_Location = 'C:/Documents'
# Set Data Location:
data = Git_Location + 'Data.csv'
# load the data
df = pd.read_csv(data,low_memory=False,thousands=',', encoding='latin-1')
df = df[['CODE','Description']] #select only these columns
df = df.rename(index=float, columns={"CODE": "label", "Description": "text"})
#Convert label to float so you don't need to encode for processing later on
df['label']=df['label'].str.replace('-', '',regex=True, case = False).str.strip()
df['label'].astype('float64', raise_on_error = True)
# drop any labels with count LT 500 to build a strong model and make our testing run faster -- we will get more data later
df = df.groupby('label').filter(lambda x : len(x)>500)
#split data into testing and training
train_x, valid_x, train_y, valid_y = model_selection.train_test_split(df.text, df.label,test_size=0.33, random_state=6,stratify=df.label)
# Other examples online use the following data types... we will do the same to remain consistent
train_y_npar = pd.Series(train_y).values
train_x_list = pd.Series.tolist(train_x)
valid_x_list = pd.Series.tolist(valid_x)
# cast validation datasets to dataframes to allow to merging later on
valid_x_df = pd.DataFrame(valid_x)
valid_y_df = pd.DataFrame(valid_y)
# Extracting features from data
count_vect = CountVectorizer()
X_train_counts = count_vect.fit_transform(train_x_list)
X_test_counts = count_vect.transform(valid_x_list)
# Define the model training and validation function
def TV_model(classifier, feature_vector_train, label, feature_vector_valid, valid_y, valid_x, is_neural_net=False):
# fit the training dataset on the classifier
classifier.fit(feature_vector_train, label)
# predict the top n labels on validation dataset
n = 5
#classifier.probability = True
probas = classifier.predict_proba(feature_vector_valid)
predictions = classifier.predict(feature_vector_valid)
#Identify the indexes of the top predictions
top_n_predictions = np.argsort(probas, axis = 1)[:,-n:]
#then find the associated SOC code for each prediction
top_class = classifier.classes_[top_n_predictions]
#cast to a new dataframe
top_class_df = pd.DataFrame(data=top_class)
#merge it up with the validation labels and descriptions
results = pd.merge(valid_y, valid_x, left_index=True, right_index=True)
results = pd.merge(results, top_class_df, left_index=True, right_index=True)
top5_conditions = [
(results.iloc[:,0] == results[0]),
(results.iloc[:,0] == results[1]),
(results.iloc[:,0] == results[2]),
(results.iloc[:,0] == results[3]),
(results.iloc[:,0] == results[4])]
top5_choices = [1, 1, 1, 1, 1]
#Top 1 Result
#top1_conditions = [(results['0_x'] == results[4])]
top1_conditions = [(results.iloc[:,0] == results[4])]
top1_choices = [1]
# Create the success columns
results['Top 5 Successes'] = np.select(top5_conditions, top5_choices, default=0)
results['Top 1 Successes'] = np.select(top1_conditions, top1_choices, default=0)
print("Are Top 5 Results greater than Top 1 Result?: ", (sum(results['Top 5 Successes'])/results.shape[0])>(metrics.accuracy_score(valid_y, predictions)))
print("Are Top 1 Results equal from predict() and predict_proba()?: ", (sum(results['Top 1 Successes'])/results.shape[0])==(metrics.accuracy_score(valid_y, predictions)))
print(" ")
print("Details: ")
print("Top 5 Accuracy Rate (predict_proba)= ", sum(results['Top 5 Successes'])/results.shape[0])
print("Top 1 Accuracy Rate (predict_proba)= ", sum(results['Top 1 Successes'])/results.shape[0])
print("Top 1 Accuracy Rate = (predict)=", metrics.accuracy_score(valid_y, predictions))
使用scikit learn内置于20个新闻组数据集的输出示例(这是我的目标):
注意:我在另一个数据集上运行了这个精确的代码,并能够生成这些结果,这告诉我函数及其依赖项可以工作,因此问题一定是在数据中。
Are Top 5 Results greater than Top 1 Result?: True
Are Top 1 Results equal from predict() and predict_proba()?: True
细节:
Top 5 Accuracy Rate (predict_proba)= 0.9583112055231015
Top 1 Accuracy Rate (predict_proba)= 0.8069569835369091
Top 1 Accuracy Rate = (predict)= 0.8069569835369091
现在运行我的数据:
TV_model(LogisticRegression(), X_train_counts, train_y_npar, X_test_counts, valid_y_df, valid_x_df)
输出:
Are Top 5 Results greater than Top 1 Result?: False
Are Top 1 Results equal from predict() and predict_proba()?: False
细节:
-
前5名准确率(预测概率)=0.6581632653061225
-
前1名准确率(预测概率)=0.2010204081632653
-
前1名准确率=(预测)=0.8091187478734263