Fix ReferentialIntegrity NaN handling

[frances-h]

• SDMetrics version: diagnostic_report_updates
• Python version:
• Operating System:

Error Description

Currently, the ReferentialIntegrity metric will count NaN foreign key values as not referencing a parent unless the parent keys also include a NaN. The logic should be updated to better handle foreign keys being null:

1. If the real foreign keys contain NaN values, the metric should ignore NaN values in the synthetic data. For example:

>>> parent_keys = ['a', 'b', 'c']
>>> foreign_keys = ['a', 'a', 'b', 'c', NaN]
>>> ReferentialIntegrity.compute_breakdown(
            real_data=(parent_keys, foreign_keys),
            synthetic_data=(parent_keys, foreign_keys))
 {'score':  1.0}

2. If the real foreign keys DO NOT contain NaN values, the metric should treat NaN values as a regular foreign key value (current functionality).

>>> parent_keys = ['a', 'b', 'c']
>>> foreign_keys =  ['a', 'a', 'b', 'c', 'a']
>>> synth_foreign_keys = ['a', 'a', 'b', 'c', NaN]
>>> ReferentialIntegrity.compute_breakdown(
            real_data=(parent_keys, foreign_keys),
            synthetic_data=(parent_keys, synth_foreign_keys))
 {'score':  0.8}

Steps to reproduce

import pandas as pd
import numpy as np

from sdmetrics.column_pairs import ReferentialIntegrity

real_primary_keys = ['a', 'b', 'c']
real_foreign_keys = ['a', 'a', 'a', 'b', 'c', np.nan]
synthetic_primary_keys= ['id1', 'id2', 'id3']
synthetic_foreign_keys = ['id1', 'id2', 'id2', 'id3', np.nan, np.nan]

ReferentialIntegrity.compute(
    real_data=(real_primary_keys, real_foreign_keys),
    synthetic_data=(synthetic_primary_keys, synthetic_foreign_keys))