每次运行K近邻分类器代码准确率不同，求原因解析

Hey Ashish, I totally get why this is confusing—let's walk through exactly what's happening here.

The Root Cause: Random Train-Test Splitting #

The main issue is how train_test_split works by default. Every time you run your code, this function randomly shuffles your dataset before splitting it into training and test sets. That means:

• Your X_train/y_train and X_test/y_test are different every run
• Since KNN relies entirely on the training data to find nearest neighbors, changing the training set changes which neighbors get picked for predictions
• This leads to different prediction results, and thus different accuracy scores

Quick Fix: Fix the Random Seed #

To make your splits (and therefore your accuracy) consistent every time you run the code, just add the random_state parameter to train_test_split. Pick any integer (like 42, a common choice for reproducibility) and set it here—this locks in the random shuffle pattern.

Corrected Code #

I also fixed a couple of small issues in your original code (like missing the import for train_test_split and unused imports):

# Removed unused imports (scipy, numpy) since you aren't using them here
from sklearn import datasets
from sklearn.model_selection import train_test_split  # Added missing import
from sklearn.neighbors import KNeighborsClassifier
from sklearn.metrics import accuracy_score

iris = datasets.load_iris()
X = iris.data
y = iris.target

# Added random_state to fix the split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=42)

my_classifier = KNeighborsClassifier()
my_classifier.fit(X_train, y_train)
predictions = my_classifier.predict(X_test)

print(predictions)
print(accuracy_score(y_test, predictions))

Now every time you run this, your training/test sets will be identical, so your accuracy score will stay the same.

Bonus Note #

If you want to make sure your model's performance is robust (not just tied to one specific train/test split), you can use techniques like cross-validation (e.g., cross_val_score from sklearn) to average accuracy across multiple splits. But for reproducibility in single runs, fixing random_state is the way to go.

内容的提问来源于stack exchange，提问作者Ashish.k