scikit cosine_similarity 对比 pairwise_distances

Question

Scikit-learn 的 sklearn.metrics.pairwise.cosine_similarity 和 sklearn.metrics.pairwise.pairwise_distances(.. metric="cosine") 有什么区别？

from sklearn.feature_extraction.text import TfidfVectorizer

documents = (
    "Macbook Pro 15' Silver Gray with Nvidia GPU",
    "Macbook GPU"    
)

tfidf_vectorizer = TfidfVectorizer()
tfidf_matrix = tfidf_vectorizer.fit_transform(documents)

from sklearn.metrics.pairwise import cosine_similarity
print(cosine_similarity(tfidf_matrix[0:1], tfidf_matrix)[0,1])

0.37997836

from sklearn.metrics.pairwise import pairwise_distances
print(pairwise_distances(tfidf_matrix[0:1], tfidf_matrix, metric='cosine')[0,1])

0.62002164

为什么这些不同？

Answer 1

来自源代码documentation:

Cosine distance is defined as 1.0 minus the cosine similarity.

所以你的结果是有道理的。

Answer 2

成对距离提供两个数组之间的距离 array.so 更多成对距离意味着更少 similarity.while 余弦相似度是 1-pairwise_distance 因此更多余弦相似度意味着两个数组之间更多相似度。