[BUG] SAR needs to be modified due to a breaking change in scipy #1954
Description
Description
With scipy 1.10.1, the item similarity matrix is a dense matrix
print(type(model.item_similarity))
print(type(model.user_affinity))
print(type(model.item_similarity) == np.ndarray)
print(type(model.item_similarity) == scipy.sparse._csr.csr_matrix)
print(model.item_similarity.shape)
print(model.item_similarity)
<class 'numpy.ndarray'>
<class 'scipy.sparse._csr.csr_matrix'>
True
False
(1646, 1646)
[[1. 0.10650888 0.03076923 ... 0. 0. 0. ]
[0.10650888 1. 0.15104167 ... 0. 0.00729927 0.00729927]
[0.03076923 0.15104167 1. ... 0. 0. 0.01190476]
...
[0. 0. 0. ... 1. 0. 0. ]
[0. 0.00729927 0. ... 0. 1. 0. ]
[0. 0.00729927 0.01190476 ... 0. 0. 1. ]]
but with scipy 1.11.1 the item similarity matrix is sparse
print(type(model.item_similarity))
print(type(model.user_affinity))
type(model.item_similarity) == np.ndarray
type(model.item_similarity) == scipy.sparse._csr.csr_matrix
print(model.item_similarity.shape)
<class 'numpy.ndarray'>
<class 'scipy.sparse._csr.csr_matrix'>
()
In which platform does it happen?
Related to #1951
How do we replicate the issue?
Expected behavior (i.e. solution)
Other Comments
We found that the issue was that during a division in Jaccard, scipy change the type. We talked to the authors of scipy and they told us that they did a breaking change in 1.11.0 scipy/scipy#18796 (comment)