Compute the similarity between two lists


I'd like to compute the similarity between two lists of various lengths.


listA = ['apple', 'orange', 'apple', 'apple', 'banana', 'orange'] # (length = 6)
listB = ['apple', 'orange', 'grapefruit', 'apple'] # (length = 4)

as you can see, a single item can appear multiple times in a list, and the lengths are of different sizes.

I've already thought of comparing the frequencies of each item, but that does not encompass the size of each list (a list that is simply twice another list should be similar, but not perfectly similar)


listA = ['apple', 'apple', 'orange', 'orange']
listB = ['apple', 'orange']
similarity(listA, listB) # should NOT equal 1

So I basically want to encompass the size of the lists, and the distribution of items in the list.

Any ideas?

2/6/2013 8:19:22 AM

Accepted Answer

Use collections.Counter() perhaps; those are multi-sets, or bags, in datatype parlance:

from collections import Counter

counterA = Counter(listA)
counterB = Counter(listB)

Now you can compare these by entries or frequencies:

>>> counterA
Counter({'apple': 3, 'orange': 2, 'banana': 1})
>>> counterB
Counter({'apple': 2, 'orange': 1, 'grapefruit': 1})
>>> counterA - counterB
Counter({'orange': 1, 'apple': 1, 'banana': 1})
>>> counterB - counterA
Counter({'grapefruit': 1})

You can calculate their cosine similarity using:

import math

def counter_cosine_similarity(c1, c2):
    terms = set(c1).union(c2)
    dotprod = sum(c1.get(k, 0) * c2.get(k, 0) for k in terms)
    magA = math.sqrt(sum(c1.get(k, 0)**2 for k in terms))
    magB = math.sqrt(sum(c2.get(k, 0)**2 for k in terms))
    return dotprod / (magA * magB)

Which gives:

>>> counter_cosine_similarity(counterA, counterB)

The closer to 1 that value, the more similar the two lists are.

The cosine similarity is one score you can calculate. If you care about the length of the list, you can calculate another; if you keep that score between 0.0 and 1.0 as well you can multiply the two values for a final score between -1.0 and 1.0.

For example, to take relative lengths into account you could use:

def length_similarity(c1, c2):
    lenc1 = sum(c1.itervalues())
    lenc2 = sum(c2.itervalues())
    return min(lenc1, lenc2) / float(max(lenc1, lenc2))

and then combine into a function that takes the lists as inputs:

def similarity_score(l1, l2):
    c1, c2 = Counter(l1), Counter(l2)
    return length_similarity(c1, c2) * counter_cosine_similarity(c1, c2)  

For your two example lists, that results in:

>>> similarity_score(['apple', 'orange', 'apple', 'apple', 'banana', 'orange'], ['apple', 'orange', 'grapefruit', 'apple'])
>>> similarity_score(['apple', 'apple', 'orange', 'orange'], ['apple', 'orange'])

You can mix in other metrics as needed.

2/6/2013 4:44:29 PM

From a theoretical point of view : I recommend you look up cosine similarity

You may have to modify to fit your scheme, but the idea of cosine similarity is great.

Licensed under: CC-BY-SA with attribution
Not affiliated with: Stack Overflow