为什么在 python 2.7 中比 3.6 有更好的性能（剪影分数）？

Question

I read a lot on SoF about the difference in speed between Python 2.7 and 3.6. but my question is more about performance between the two versions.

I used for document clustering: TF-IDF + KMeans and score silhouette to evaluate the homogeneity of my clusters.

By switching from Python 3.6 to Python 2.7, my silhouette score has increased by +0.20!

**Would someone have an explanation? ** Thanks!

code :

tfidf = TfidfVectorizer(
    stop_words=my_stopwords_str, 
    max_df=0.95, 
    min_df=5, 
    token_pattern=r'\w{3,}',
    max_features=20)

tfidf.fit(data_final.all_text)
data_vect = tfidf.transform(data_final.all_text)

num_clusters = 15

kmeans = KMeans(n_clusters=num_clusters, init='k-means++', 
max_iter=300).fit(data_vect_lsa)
kmeans_predict = KMeans(n_clusters=num_clusters, init='k-means++', max_iter=300).fit_predict(data_vect_lsa)


silhouette_score(data_vect, labels = kmeans_predict, metric='euclidean')

The output for Python 2.7 is :

0.58234789374593758

The output for Python 3.6 is :

0.37524101598378656    

Answer 1

Try again. A single sample is not enough.

K-means begins with a random setting, and may find a local optimum only.

It's fairly common to see different results when running it multiple times.