Now the distances in the MLE estimator are normalized to the k-th
neighbor, as they should. Before this change, distances were always
normalized to the last vector in the collection of distances, which in
our particular setup was 100. Therefore estimates for k=100 (the ones
used in the paper) are OK, but estimates with a different k are not.
The figure below shows the distribution of LID scores computed for the GLOVE dataset with k=10 and k=100.
The two distributions are very similar
maumueller
commented
3 years ago
Now the distances in the MLE estimator are normalized to the k-th neighbor, as they should. Before this change, distances were always normalized to the last vector in the collection of distances, which in our particular setup was 100. Therefore estimates for k=100 (the ones used in the paper) are OK, but estimates with a different k are not.
The figure below shows the distribution of LID scores computed for the GLOVE dataset with k=10 and k=100. The two distributions are very similar