Index.intersection bug?

[tom-alcorn]

I ran into an odd behaviour when taking the intersection of two Indexes. Specifically,

left = pd.Index(['A','B','A','C'])
right = pd.Index(['B','D'])
left.intersection(right)

returns

Index(['B', 'C'], dtype='object')

However, I would expect this to return

Index(['B'], dtype='object')

If Index(['B', 'C'], dtype='object') is the intended behaviour, can someone explain the rationale behind it?

[jreback]

can I show pandas version

[tom-alcorn]

>>> pd.version.version
'0.14.1'

[jreback]

yep this is a non-tested case of non-monotinic non-unique indexers

which results in index.take([1,-1]) in this case (which incorrectly works; the -1 is an indexer that should be filtered out).

care to do a pull-request? (last case in core/index.py.Index.intersection)

[tom-alcorn]

This looks like the problem:

try:
    indexer = self.get_indexer(other.values)
    indexer = indexer.take((indexer != -1).nonzero()[0])
except:
    # duplicates
    indexer = self.get_indexer_non_unique(other.values)[0].unique()

Do you think this would make more sense?

try:
    indexer = self.get_indexer_non_unique(other.values)[0].unique()
    indexer = self.get_indexer(other.values)
    indexer = indexer.take((indexer != -1).nonzero()[0])
except:
    raise

[jreback]

no much simpler

before the take do

indexer = indexer[indexer != -1]

-1 is the get_indexer placeholder for missing so needs to be taken out

[jreback]

closed by #8374