基于Numpy的余弦相似度函数：最相似向量索引求解求助

Hey there! Let's walk through what's off with your current code and get it working properly—totally get being stuck as a newbie, this stuff takes time to wrap your head around.

First, let's recap what your function is supposed to do: take a vector x and a list of vectors v_list, calculate the cosine similarity between x and every vector in v_list, then return the index of the vector with the highest similarity.

What's Wrong With the Original Code? #

Your code has a few key issues that are keeping it from working as intended:

• Dimension Mismatch & Incorrect Dot Product: If v_list is a regular Python list of vectors, np.dot(x, v_list) will either throw an error or compute something you don't want. Numpy needs inputs to be properly shaped arrays to calculate dot products per vector.
• Incorrect Norm Calculation: np.linalg.norm(v_list) computes the norm of the entire list as a single tensor, not the norm of each individual vector in the list. You need to specify which axis to compute the norm along.
• No Index Return: Your current code returns similarity values instead of the index of the most similar vector.

The Fixed Code #

Here's a corrected version with explanations for each step:

import numpy as np

def most_similar(x, v_list):
    # Convert inputs to numpy arrays to handle vector operations smoothly
    x = np.asarray(x)
    v_array = np.asarray(v_list)
    
    # Calculate dot product between x and every vector in v_list
    # This gives an array where each element is the dot product of x with one vector from v_list
    dot_products = np.dot(v_array, x)
    
    # Compute the L2 norm of x (a single scalar value)
    norm_x = np.linalg.norm(x)
    
    # Compute the L2 norm of each vector in v_list
    # axis=1 tells numpy to calculate the norm for each row (each individual vector)
    norms_v = np.linalg.norm(v_array, axis=1)
    
    # Calculate cosine similarity for each vector
    # Numpy's broadcasting handles multiplying the scalar norm_x with each element in norms_v
    similarities = dot_products / (norm_x * norms_v)
    
    # Find the index of the highest similarity value
    return np.argmax(similarities)

Let's Test It Out #

Let's run a quick example to make sure it works:

# Test case: x is [1,2,3], v_list has a vector that's exactly half of x (highest similarity)
x = [1, 2, 3]
v_list = [[4,5,6], [0.5,1,1.5], [7,8,9]]
print(most_similar(x, v_list))  # Output: 1 (since the second vector is a scaled version of x, similarity = 1)

Key Notes to Remember #

• Consistent Dimensions: Make sure every vector in v_list has the same number of elements as x—otherwise numpy will throw an error.
• Cosine Similarity Basics: Cosine similarity ranges from -1 (completely opposite) to 1 (identical direction). Using np.argmax makes sense here because we want the highest similarity score.

内容的提问来源于stack exchange，提问作者Seth