Appending pandas DataFrame with MultiIndex with da

问题:

Base scenario

For a recommendation service I am training a matrix factorization model (LightFM) on a set of user-item interactions. For the matrix factorization model to yield the best results, I need to map my user and item IDs to a continuous range of integer IDs starting at 0.

I'm using a pandas DataFrame in the process, and I have found a MultiIndex to be extremely convenient to create this mapping, like so:

ratings = [{'user_id': 1, 'item_id': 1, 'rating': 1.0},
           {'user_id': 1, 'item_id': 3, 'rating': 1.0},
           {'user_id': 3, 'item_id': 1, 'rating': 1.0},
           {'user_id': 3, 'item_id': 3, 'rating': 1.0}]

df = pd.DataFrame(ratings, columns=['user_id', 'item_id', 'rating'])
df = df.set_index(['user_id', 'item_id'])
df

Out:
                 rating
user_id item_id 
1       1        1.0
1       3        1.0
3       1        1.0
3       1        1.0

And then allows me to get the continuous maps like so

df.index.labels[0]    # For users

Out:
FrozenNDArray([0, 0, 1, 1], dtype='int8')

df.index.labels[1]    # For items

Out:
FrozenNDArray([0, 1, 0, 1], dtype='int8')

Afterwards, I can map them back using df.index.levels[0].get_loc method. Great!

Extension

But, now I'm trying to streamline my model training process, ideally by training it incrementally on new data, preserving the old ID mappings. Something like:

new_ratings = [{'user_id': 2, 'item_id': 1, 'rating': 1.0},
               {'user_id': 2, 'item_id': 2, 'rating': 1.0}]

df2 = pd.DataFrame(new_ratings, columns=['user_id', 'item_id', 'rating'])
df2 = df2.set_index(['user_id', 'item_id'])
df2

Out:
                 rating
user_id item_id 
2       1        1.0
2       2        1.0

Then, simply appending the new ratings to the old DataFrame

df3 = df.append(df2)
df3

Out:
                 rating
user_id item_id 
1       1        1.0
1       3        1.0
3       1        1.0
3       3        1.0
2       1        1.0
2       2        1.0

Looks good, but

df3.index.labels[0]    # For users

Out:
FrozenNDArray([0, 0, 2, 2, 1, 1], dtype='int8')

df3.index.labels[1]    # For items

Out:
FrozenNDArray([0, 2, 0, 2, 0, 1], dtype='int8')

I added user_id=2 and item_id=2 in the later DataFrame on purpose, to illustrate where it goes wrong for me. In df3, labels 3 (for both user and item), have moved from integer position 1 to 2. So the mapping is no longer the same. What I'm looking for is [0, 0, 1, 1, 2, 2] and [0, 1, 0, 1, 0, 2] for user and item mappings respectively.

This is probably because of ordering in pandas Index objects, and I'm unsure if what I want is at all possible using a MultiIndex strategy. Looking for help on how most to effectively tackle this problem :)

Some notes:

• I find using DataFrames convenient for several reasons, but I use the MultiIndex purely for the ID mappings. Alternatives without MultiIndex are completely acceptable.
• I cannot guarantee that new user_id and item_id entries in new ratings are larger than any values in the old dataset, hence my example of adding id 2 when [1, 3] were present.
• For my incremental training approach, I will need to store my ID maps somewhere. If I only load new ratings partially, I will have to store the old DataFrame and ID maps somewhere. Would be great if it could all be in one place, like it would be with an index, but columns work too.
• EDIT: An additional requirement is to allow for row re-ordering of the original DataFrame, as might happen when duplicate ratings exist, and I want to keep the most recent one.

Solution (credits to @jpp for original)

I've made a modification to @jpp's answer to satisfy the additional requirement I've added later (tagged with EDIT). This also truly satisfies the original question as posed in the title, since it preserves the old index integer positions, regardless of rows being reordered for whatever reason. I've also wrapped things into functions:

from itertools import chain
from toolz import unique


def expand_index(source, target, index_cols=['user_id', 'item_id']):

    # Elevate index to series, keeping source with index
    temp = source.reset_index()
    target = target.reset_index()

    # Convert columns to categorical, using the source index and target columns
    for col in index_cols:
        i = source.index.names.index(col)
        col_cats = list(unique(chain(source.index.levels[i], target[col])))

        temp[col] = pd.Categorical(temp[col], categories=col_cats)
        target[col] = pd.Categorical(target[col], categories=col_cats)

    # Convert series back to index
    source = temp.set_index(index_cols)
    target = target.set_index(index_cols)

    return source, target


def concat_expand_index(old, new):
    old, new = expand_index(old, new)
    return pd.concat([old, new])


df3 = concat_expand_index(df, df2)

The result:

df3.index.labels[0]    # For users

Out:
FrozenNDArray([0, 0, 1, 1, 2, 2], dtype='int8')

df3.index.labels[1]    # For items

Out:
FrozenNDArray([0, 1, 0, 1, 0, 2], dtype='int8')

回答1:

Forcing alignment of index labels after concatenation does not appear straightforward and, if there is a solution, it is poorly documented.

One option which may appeal to you is Categorical Data. With some careful manipulation, this can achieve the same purpose: each unique index value within a level has a one-to-one mapping to an integer, and this mapping persists even after concatenation with other dataframes.

from itertools import chain
from toolz import unique

# elevate index to series
df = df.reset_index()
df2 = df2.reset_index()

# define columns for reindexing
index_cols = ['user_id', 'item_id']

# convert to categorical with merged categories
for col in index_cols:
    col_cats = list(unique(chain(df[col], df2[col])))
    df[col] = pd.Categorical(df[col], categories=col_cats)
    df2[col] = pd.Categorical(df2[col], categories=col_cats)

# convert series back to index
df = df.set_index(index_cols)
df2 = df2.set_index(index_cols)

I use toolz.unique to return an ordered unique list, but if you don't have access to this library, you can use the identical unique_everseen recipe from the itertool docs.

Now let's have a look at the category codes underlying the 0th index level:

for data in [df, df2]:
    print(data.index.get_level_values(0).codes.tolist())

[0, 0, 1, 1]
[2, 2]

Then perform our concatenation:

df3 = pd.concat([df, df2])

Finally, check that categorical codes are aligned:

print(df3.index.get_level_values(0).codes.tolist())
[0, 0, 1, 1, 2, 2]

For each index level, note we must take the union of all index values across dataframes to form col_cats, otherwise the concatenation will fail.

回答2:

I think the use of MultiIndex overcomplicates this objective:

I need to map my user and item IDs to a continuous range of integer IDs starting at 0.

This solution falls in to the below category:

Alternatives without MultiIndex are completely acceptable.

---

def add_mapping(df, df2, df3, column_name='user_id'):

    initial = df.loc[:, column_name].unique()
    new = df2.loc[~df2.loc[:, column_name].isin(initial), column_name].unique()
    maps = np.arange(len(initial))
    mapping = dict(zip(initial, maps))
    maps = np.append(maps, np.arange(np.max(maps)+1, np.max(maps)+1+len(new)))
    total = np.append(initial, new)
    mapping = dict(zip(total, maps))

    df3[column_name+'_map'] = df3.loc[:, column_name].map(mapping) 

    return df3

add_mapping(df, df2, df3, column_name='item_id')
add_mapping(df, df2, df3, column_name='user_id')

 user_id    item_id rating  item_id_map user_id_map
0   1          1    1.0         0           0
1   1          3    1.0         1           0
2   3          1    1.0         0           1
3   3          3    1.0         1           1
0   2          1    1.0         0           2
1   2          2    1.0         2           2

---

Explaination

This is how to maintain a mapping for the user_id values. Same holds for the item_id values as well.

These are the initial user_id values (unique):

initial_users = df['user_id'].unique()
# initial_users = array([1, 3])

user_map maintains a mapping for user_id values, as per your requirement:

user_id_maps = np.arange(len(initial_users))
# user_id_maps = array([0, 1])

user_map = dict(zip(initial_users, user_id_maps))
# user_map = {1: 0, 3: 1}

These are the new user_id values you got from df2 - ones that you didn't see in df:

new_users = df2[~df2['user_id'].isin(initial_users)]['user_id'].unique()
# new_users = array([2])

Now we update user_map for the total user base with the new users:

user_id_maps = np.append(user_id_maps, np.arange(np.max(user_id_maps)+1, np.max(user_id_maps)+1+len(new_users)))
# array([0, 1, 2])
total_users = np.append(initial_users, new_users)
# array([1, 3, 2])

user_map = dict(zip(total_users, user_id_maps))
# user_map = {1: 0, 2: 2, 3: 1}

Then, just map the values from user_map to df['user_id']:

df3['user_map'] = df3['user_id'].map(user_map)

user_id item_id rating  user_map
0   1   1       1.0          0
1   1   3       1.0          0
2   3   1       1.0          1
3   3   3       1.0          1
0   2   1       1.0          2
1   2   2       1.0          2