Dask.dataframe or Alternative: Scalable way of dropping rows of low frequency items

Question

I am looking for a way to remove rows from a dataframe that contain low frequency items. I adapted the following snippet from this post:

import pandas as pd
import numpy as np

df = pd.DataFrame(np.random.randint(0, high=9, size=(100,2)),
         columns = ['A', 'B'])

threshold = 10 # Anything that occurs less than this will be removed.
value_counts = df.stack().value_counts() # Entire DataFrame 
to_remove = value_counts[value_counts <= threshold].index
df.replace(to_remove, np.nan, inplace=True)

The problem is, that this code does not scale, it seems.

The line to_remove = value_counts[value_counts <= threshold].index has now been running for several hours for my data (2 GB compressed HDFStore). I therefore need a better solution. Ideally out-of-core. I suspect dask.dataframe is suitable, but I fail to express the above code in terms of dask. The key functions stack and replace are absent from dask.dataframe.

I tried the following (works in normal pandas) to work around the lack of these two functions:

value_countss = [df[col].value_counts() for col in df.columns]

infrequent_itemss = [value_counts[value_counts < 3] for value_counts in value_countss]

rows_to_drop = set(i for indices in [df.loc[df[col].isin(infrequent_items.keys())].index.values  for col, infrequent_items in zip(df.columns, infrequent_itemss)] for i in indices)

df.drop(rows_to_drop)

That does not actually work with dask though. It errors at infrequent_items.keys().

Even if it did work, given that this is the opposite of elegant, I suspect there must be a better way.

Can you suggest something?

To use with dask, most likely you will want to try [`map_partitions`](http://docs.dask.org/en/latest/dataframe-api.html#dask.dataframe.DataFrame.map_partitions) — mdurant, Dec 05 '18 at 19:16
Are "partitions" and "chunks" the same thing there? In other context functions have a `chunksize` parameter. — lo tolmencre, Dec 05 '18 at 19:22

score 1 · Answer 1 · answered Dec 04 '18 at 07:19

1

Not sure if this will help you out, but it's too big for a comment:

df = pd.DataFrame(np.random.randint(0, high=20, size=(30,2)), columns = ['A', 'B'])
unique, counts = np.unique(df.values.ravel(), return_counts=True)
d = dict(zip(unique, counts))
threshold = 10
to_remove = [k for k, v in d.items() if v < threshold]
df.replace(to_remove, np.nan, inplace=True)

See:

How to count the occurrence of certain item in an ndarray in Python?

how to count occurrence of each unique value in pandas

Toy problem showed a 40x speedup from 400 us to 10 us in the step you mentioned.

answered Dec 04 '18 at 07:19

Evan

2,121
14
27

Unfortunatly this still takes ages in the last line (`df.replace(to_remove, np.nan, inplace=True)`). I killed the process after a few hours. Any idea what is so time consuming about `DataFrame.replace`? my dataframe is categorical dtype with strings. – lo tolmencre Dec 04 '18 at 14:56
Well, progress! check this out? https://stackoverflow.com/questions/41985566/pandas-replace-dictionary-slowness Using `.map` or `.applymap` might be faster; apparently `df.replace()` is known to be slow but I am not sure why. – Evan Dec 04 '18 at 15:24

lo tolmencre · Accepted Answer · 2018-12-06T01:04:53.223

The following code, which incorporates Evan's improvement, solves my issue:

unique, counts = np.unique(df.values.ravel(), return_counts=True)
d = dict(zip(unique, counts))

to_remove = {k for k, v in d.items() if v < threshold}

mask = df.isin(to_remove)

column_mask = (~mask).all(axis=1)

df = df[column_mask]

demo:

def filter_low_frequency(df, threshold=4):
    unique, counts = np.unique(df.values.ravel(), return_counts=True)
    d = dict(zip(unique, counts))
    to_remove = {k for k, v in d.items() if v < threshold}
    mask = df.isin(to_remove)
    column_mask = (~mask).all(axis=1)
    df = df[column_mask]
    return df


df = pd.DataFrame(np.random.randint(0, high=20, size=(10,10)))
print(df)
print(df.stack().value_counts())
df = filter_low_frequency(df)
print(df)

    0   1   2   3   4   5   6   7   8   9
0   3  17  11  13   8   8  15  14   7   8
1   2  14  11   3  16  10  19  19  14   4
2   8  13  13  17   3  13  17  18   5  18
3   7   8  14   9  15  12   0  15   2  19
4   6  12  13  11  16   6  19  16   2  17
5   2   1   2  17   1   3  12  10   2  16
6   0  19   9   4  15   3   3   3   4   0
7  18   8  15   9   1  18  15  17   9   0
8  17  15   9  11  13   9  11   4  19   8
9  13   6   7   8   8  10   0   3  16  13
8     9
3     8
13    8
17    7
15    7
19    6
2     6
9     6
11    5
16    5
0     5
18    4
4     4
14    4
10    3
12    3
7     3
6     3
1     3
5     1
dtype: int64
    0   1  2   3   4  5   6  7   8  9
6   0  19  9   4  15  3   3  3   4  0
8  17  15  9  11  13  9  11  4  19  8

Dask.dataframe or Alternative: Scalable way of dropping rows of low frequency items

2 Answers2

Linked