This PR ensures that in an edge case where all examples in a dataset have identical features, they have the expected outlier- and near-duplicate scores.

It slightly increases the minimum median values over the machine epsilon to avoid some numerically unstable operations with floating point round-off errors.

MWE

A small dataset of 20 examples with identical features will only have (exact) duplicates but no outliers.
The scores for these issue types will be 0.0 and 1.0 for near-duplicates and outliers, respectively.

import numpy as np
from cleanlab import Datalab

# Creating a dataset with all identical examples
X = np.full((20, 5), fill_value=np.random.rand(5))
lab = Datalab(data={"X": X})
lab.find_issues(features=X)

lab.report(show_all_issues=True)

# ...
# ------------------ near_duplicate issues -------------------
# ...
# Overall dataset quality in terms of this issue: 0.0000
# ...
# ---------------------- outlier issues ----------------------
# Overall dataset quality in terms of this issue: 1.0000
# ...

Not covered in this PR

• Consider label issues for such a dataset.
• Special considerations for datasets with a single, large set of exact duplicates (much larger than what may be found in a typical knn-graph).
  • The near-duplicate sets belonging to two exact duplicate examples should be equivalent (up the the indices of the exact duplicates themselves). The transitive relation among these sets must be addressed in a separate PR.
• Cases in modules like cleanlab.object_detection.filter.py where machine epsilons appear in denominators in divisions.
  

  cleanlab/cleanlab/object_detection/filter.py

  Lines 305 to 307 in 589edd6

  	eps = np.finfo(np.float32).eps
  	recalls = true_positives / np.maximum(num_gts[:, np.newaxis], eps)
  	precisions = true_positives / np.maximum((true_positives + false_positives), eps)

• Other tiny values appearing in such places like
  

  cleanlab/cleanlab/internal/constants.py

  Line 23 in 61e9d6d

  TINY_VALUE = 1e-100 # very tiny value for clipping

  
  

  cleanlab/cleanlab/internal/latent_algebra.py

  Line 127 in 61e9d6d

  inverse_noise_matrix = joint.T / np.clip(ps, a_min=TINY_VALUE, a_max=None)

Related Issues / PRs

This PR partially addresses #1055.