Calculate intersection over union (Jaccard's index) in pandas dataframe

user_2479947 0
Viewed 424

I have a dataframe like:

animal    ids
cat       1,3,4
dog       1,2,4
hamster   5        
dolphin   3,5

The dataframe is quite big, with over 80 thousand rows, and ids column may contain easily over thousands, even 10 thousands comma separated id. Ids in a given row would be unique in the comma separated string.

I would like to construct a dataframe which calculated Jaccard's index, i.e. intersection of each items in animal column with each other in ids column over union.

So if we look at cat and dog, the union is 2 (ids 1 and 4), and union is 4 (ids 1, 2, 3, 4), hence the Jaccard's index is 2/4 = 0.5. It would be great to have the dataset in this format:

            cat        dog        hamster    dolphin
cat         1          0.5        0          0.25
dog         0.5        1          0          0
hamster     0          0          1          0.5
dolphin     0.25       0          0.5        1

which means using the row index as the name of the animal, so that I can find related jaccard's index quickly like:

cat_dog_ji = df_new['cat']['dog']
scikit-learn pandas dataframe numpy python
2 Answers

You can use str.get_dummies and some scipy tools here.

from scipy.spatial import distance

u = df["ids"].str.get_dummies(",")
j = distance.pdist(u, "jaccard")
k = df["animal"].to_numpy()
pd.DataFrame(1 - distance.squareform(j), index=k, columns=k)

          cat  dog  hamster  dolphin
cat      1.00  0.5      0.0     0.25
dog      0.50  1.0      0.0     0.00
hamster  0.00  0.0      1.0     0.50
dolphin  0.25  0.0      0.5     1.00
user_34832030

Use:

d = df.assign(key=1, ids=df['ids'].str.split(','))
d = d.merge(d, on='key', suffixes=['', '_r'])

i = [np.intersect1d(*x).size / np.union1d(*x).size for x in zip(d['ids'], d['ids_r'])]
d = pd.crosstab(d['animal'], d['animal_r'], i, aggfunc='first').rename_axis(index=None, columns=None)

Details:

Use DataFrame.assign to create a temporary column key and use Series.str.split on column ids. Then use DataFrame.merge to merge the dataframe d with itself based column key (essentially a cross join).

print(d)

     animal        ids  key animal_r      ids_r
0       cat  [1, 3, 4]    1      cat  [1, 3, 4]
1       cat  [1, 3, 4]    1      dog  [1, 2, 4]
2       cat  [1, 3, 4]    1  hamster        [5]
3       cat  [1, 3, 4]    1  dolphin     [3, 5]
4       dog  [1, 2, 4]    1      cat  [1, 3, 4]
5       dog  [1, 2, 4]    1      dog  [1, 2, 4]
6       dog  [1, 2, 4]    1  hamster        [5]
7       dog  [1, 2, 4]    1  dolphin     [3, 5]
8   hamster        [5]    1      cat  [1, 3, 4]
9   hamster        [5]    1      dog  [1, 2, 4]
10  hamster        [5]    1  hamster        [5]
11  hamster        [5]    1  dolphin     [3, 5]
12  dolphin     [3, 5]    1      cat  [1, 3, 4]
13  dolphin     [3, 5]    1      dog  [1, 2, 4]
14  dolphin     [3, 5]    1  hamster        [5]
15  dolphin     [3, 5]    1  dolphin     [3, 5]

Using np.interset1d along with np.union1d inside list comprehension to calculate the Jaccard's index.

print(i)
[1.0, 0.5, 0.0, 0.25, 0.5, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.5, 0.25, 0.0, 0.5, 1.0]

Finally we use pd.crosstab to create a simple cross tabulation to get the result in desired format:

print(d)
          cat  dog  dolphin  hamster
cat      1.00  0.5     0.25      0.0
dog      0.50  1.0     0.00      0.0
dolphin  0.25  0.0     1.00      0.5
hamster  0.00  0.0     0.50      1.0
user_128331660
Related