cosine_similarity between 2 pandas df column to get cosine distance

Question

I have a dataframe as shown below:

vector_a            vector_b
[1,2,3]             [2,5,6]
[0,2,1]             [2,9,1]
[4,7,1]             [1,7,4]

I would like to do sklearn's cosine_similarity between the columns vector_a and vector_b to get a new column called 'cosine_distance' in the same dataframe. Do note that vector_a and vector_b are pandas df columns of list.

This is what I have attempted:

df['vector_a'] = df['vector_a'].apply(lambda x: np.asarray(x))
df['vector_b'] = df['vector_b'].apply(lambda x: np.asarray(x))
df['cosine_distance'] = cosine_similarity(df['vector_a'].apply(lambda x: np.transpose(x)), 
                                          df['vector_b'].apply(lambda x: np.transpose(x)))

And I got this error:

---> 58         df['cosine_distance'] = cosine_similarity(df['vector_a'].apply(lambda x: np.transpose(x)), df['vector_b'].apply(lambda x: np.transpose(x)))

~\Anaconda3\lib\site-packages\sklearn\metrics\pairwise.py in cosine_similarity(X, Y, dense_output)
   1025     # to avoid recursive import
   1026 
-> 1027     X, Y = check_pairwise_arrays(X, Y)
   1028 
   1029     X_normalized = normalize(X, copy=True)

~\Anaconda3\lib\site-packages\sklearn\metrics\pairwise.py in check_pairwise_arrays(X, Y, precomputed, dtype)
    110     else:
    111         X = check_array(X, accept_sparse='csr', dtype=dtype,
--> 112                         estimator=estimator)
    113         Y = check_array(Y, accept_sparse='csr', dtype=dtype,
    114                         estimator=estimator)

~\Anaconda3\lib\site-packages\sklearn\utils\validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, warn_on_dtype, estimator)
    494             try:
    495                 warnings.simplefilter('error', ComplexWarning)
--> 496                 array = np.asarray(array, dtype=dtype, order=order)
    497             except ComplexWarning:
    498                 raise ValueError("Complex data not supported\n"

~\Anaconda3\lib\site-packages\numpy\core\numeric.py in asarray(a, dtype, order)
    536 
    537     """
--> 538     return array(a, dtype, copy=False, order=order)
    539 
    540 

ValueError: setting an array element with a sequence.

Thank you in advance!

score 4 · Accepted Answer · answered Jan 01 '20 at 00:35

TLDR:

df['cosine_similarity'] = df.apply(
    lambda row: cosine_similarity([row['vector_a']], [row['vector_b']])[0][0],
    axis=1)

Explanation:

cosine_similarity expects 2D np.array, or list of lists. It doesn't know how to interpret pd.Series of lists. However, even if we did convert it to list of lists, the next problem arises:
cosine_similarity returns all-vs-all similarity. So, let's limit to pairwise comparison, artificially creating second dimension (note the extra square brackets in [row['vector_a']], [row['vector_b']]), and then taking the only element of a 1x1 array (zeros at the end of cosine_similarity(...)[0][0])

cosine_similarity between 2 pandas df column to get cosine distance

1 Answers1