cjnolet
commented
2 months ago @QuotoAstro,
The pairwise distances calculate distance, not similarity. This is consistent with Sciki-learn also. As you’ve noted, you can turn the distance into a similarity by subtracting all the values from 1.
Describe the bug
When use cuml.metrics.pairwise_distances calculate cosine similarity, the program return the wrong answer.
Steps/Code to reproduce bug
As shown in the doc.
Expected behavior
Actually, for example, the expected result is 0.554 for the [2, 3] and [1, 0]. Seems all the wrong results are resulted from 1 - correct result.