Customizing rolling_apply function in Python pandas

Question

Setup

I have a DataFrame with three columns:

"Category" contains True and False, and I have done df.groupby('Category') to group by these values.
"Time" contains timestamps (measured in seconds) at which values have been recorded
"Value" contains the values themselves.

At each time instance, two values are recorded: one has category "True", and the other has category "False".

Rolling apply question

Within each category group, I want to compute a number and store it in column Result for each time. Result is the percentage of values between time t-60 and t that fall between 1 and 3.

The easiest way to accomplish this is probably to calculate the total number of values in that time interval via rolling_count, then execute rolling_apply to count only the values from that interval that fall between 1 and 3.

Here is my code so far:

groups = df.groupby(['Category'])
for key, grp in groups:
    grp = grp.reindex(grp['Time']) # reindex by time so we can count with rolling windows
    grp['total'] = pd.rolling_count(grp['Value'], window=60) # count number of values in the last 60 seconds
    grp['in_interval'] = ? ## Need to count number of values where 1<v<3 in the last 60 seconds

    grp['Result'] = grp['in_interval'] / grp['total'] # percentage of values between 1 and 3 in the last 60 seconds

What is the proper rolling_apply() call to find grp['in_interval']?

unutbu · Accepted Answer · 2013-03-19 04:08:26Z

Let's work through an example:

import pandas as pd
import numpy as np
np.random.seed(1)

def setup(regular=True):
    N = 10
    x = np.arange(N)
    a = np.arange(N)
    b = np.arange(N)

    if regular:
        timestamps = np.linspace(0, 120, N)
    else:
        timestamps = np.random.uniform(0, 120, N)

    df = pd.DataFrame({
        'Category': [True]*N + [False]*N,
        'Time': np.hstack((timestamps, timestamps)),
        'Value': np.hstack((a,b))
        })
    return df

df = setup(regular=False)
df.sort(['Category', 'Time'], inplace=True)

So the DataFrame, df, looks like this:

In [4]: df
Out[4]: 
   Category       Time  Value    Result
12    False   0.013725      2  1.000000
15    False  11.080631      5  0.500000
14    False  17.610707      4  0.333333
16    False  22.351225      6  0.250000
13    False  36.279909      3  0.400000
17    False  41.467287      7  0.333333
18    False  47.612097      8  0.285714
10    False  50.042641      0  0.250000
19    False  64.658008      9  0.125000
11    False  86.438939      1  0.333333
2      True   0.013725      2  1.000000
5      True  11.080631      5  0.500000
4      True  17.610707      4  0.333333
6      True  22.351225      6  0.250000
3      True  36.279909      3  0.400000
7      True  41.467287      7  0.333333
8      True  47.612097      8  0.285714
0      True  50.042641      0  0.250000
9      True  64.658008      9  0.125000
1      True  86.438939      1  0.333333

Now, copying @herrfz, let's define

def between(a, b):
    def between_percentage(series):
        return float(len(series[(a <= series) & (series < b)])) / float(len(series))
    return between_percentage

between(1,3) is a function which takes a Series as input and returns the fraction of its elements which lie in the half-open interval [1,3). For example,

In [9]: series = pd.Series([1,2,3,4,5])

In [10]: between(1,3)(series)
Out[10]: 0.4

Now we are going to take our DataFrame, df, and group by Category:

df.groupby(['Category'])

For each group in the groupby object, we will want to apply a function:

df['Result'] = df.groupby(['Category']).apply(toeach_category)

The function, toeach_category, will take a (sub)DataFrame as input, and return a DataFrame as output. The entire result will be assigned to a new column of df called Result.

Now what exactly must toeach_category do? If we write toeach_category like this:

def toeach_category(subf):
    print(subf)

then we see each subf is a DataFrame such as this one (when Category is False):

   Category       Time  Value    Result
12    False   0.013725      2  1.000000
15    False  11.080631      5  0.500000
14    False  17.610707      4  0.333333
16    False  22.351225      6  0.250000
13    False  36.279909      3  0.400000
17    False  41.467287      7  0.333333
18    False  47.612097      8  0.285714
10    False  50.042641      0  0.250000
19    False  64.658008      9  0.125000
11    False  86.438939      1  0.333333

We want to take the Times column, and for each time, apply a function. That's done with applymap:

def toeach_category(subf):
    result = subf[['Time']].applymap(percentage)

The function percentage will take a time value as input, and return a value as output. The value will be the fraction of rows with values between 1 and 3. applymap is very strict: percentage can not take any other arguments.

Given a time t, we can select the Values from subf whose times are in the half-open interval (t-60, t] using the ix method:

subf.ix[(t-60 < subf['Time']) & (subf['Time'] <= t), 'Value']

And so we can find the percentage of those Values between 1 and 3 by applying between(1,3):

between(1,3)(subf.ix[(t-60 < subf['Time']) & (subf['Time'] <= t), 'Value'])

Now remember that we want a function percentage which takes t as input and returns the above expression as output:

def percentage(t):
    return between(1,3)(subf.ix[(t-60 < subf['Time']) & (subf['Time'] <= t), 'Value'])

But notice that percentage depends on subf, and we are not allowed to pass subf to percentage as an argument (again, because applymap is very strict).

So how do we get out of this jam? The solution is to define percentage inside toeach_category. Python's scoping rules say that a bare name like subf is first looked for in the Local scope, then the Enclosing scope, the the Global scope, and lastly in the Builtin scope. When percentage(t) is called, and Python encounters subf, Python first looks in the Local scope for the value of subf. Since subf is not a local variable in percentage, Python looks for it in the Enclosing scope of the function toeach_category. It finds subf there. Perfect. That is just what we need.

So now we have our function toeach_category:

def toeach_category(subf):
    def percentage(t):
        return between(1, 3)(
            subf.ix[(t - 60 < subf['Time']) & (subf['Time'] <= t), 'Value'])
    result = subf[['Time']].applymap(percentage)
    return result

Putting it all together,

import pandas as pd
import numpy as np
np.random.seed(1)


def setup(regular=True):
    N = 10
    x = np.arange(N)
    a = np.arange(N)
    b = np.arange(N)

    if regular:
        timestamps = np.linspace(0, 120, N)
    else:
        timestamps = np.random.uniform(0, 120, N)

    df = pd.DataFrame({
        'Category': [True] * N + [False] * N,
        'Time': np.hstack((timestamps, timestamps)),
        'Value': np.hstack((a, b))
    })
    return df


def between(a, b):
    def between_percentage(series):
        return float(len(series[(a <= series) & (series < b)])) / float(len(series))
    return between_percentage


def toeach_category(subf):
    def percentage(t):
        return between(1, 3)(
            subf.ix[(t - 60 < subf['Time']) & (subf['Time'] <= t), 'Value'])
    result = subf[['Time']].applymap(percentage)
    return result


df = setup(regular=False)
df.sort(['Category', 'Time'], inplace=True)
df['Result'] = df.groupby(['Category']).apply(toeach_category)
print(df)

yields

   Category       Time  Value    Result
12    False   0.013725      2  1.000000
15    False  11.080631      5  0.500000
14    False  17.610707      4  0.333333
16    False  22.351225      6  0.250000
13    False  36.279909      3  0.200000
17    False  41.467287      7  0.166667
18    False  47.612097      8  0.142857
10    False  50.042641      0  0.125000
19    False  64.658008      9  0.000000
11    False  86.438939      1  0.166667
2      True   0.013725      2  1.000000
5      True  11.080631      5  0.500000
4      True  17.610707      4  0.333333
6      True  22.351225      6  0.250000
3      True  36.279909      3  0.200000
7      True  41.467287      7  0.166667
8      True  47.612097      8  0.142857
0      True  50.042641      0  0.125000
9      True  64.658008      9  0.000000
1      True  86.438939      1  0.166667

This looks like an interesting and detailed solution. But why not use apply_rolling as the original question asks? surely it can be used? I have a very similar problem.
@Carl: pd.rolling_apply requires a fixed (integer) window size. In this problem, the window size changes with each row, since the window depends on the value in the Time column.
Thanks unutbu, much appreciated. In the case of fixed window size, how would one use it?
That question is rather broad. Please open a new question that spells out the details, or google "pandas pd.rolling_apply site: stackoverflow.com" to find lots of examples.

herrfz · Accepted Answer · 2013-03-18 21:55:56Z

2

If I understand your problem statement correctly, you could probably skip rolling count if you use it only for the sake of computing the percentage. rolling_apply takes as an argument a function that performs aggregation, i.e. a function that takes an array as input and returns a number as an output.

Having this in mind, let's first define a function:

def between_1_3_perc(x):
    # pandas Series is basically a numpy array, we can do boolean indexing
    return float(len(x[(x > 1) & (x < 3)])) / float(len(x))

Then use the function name as an argument of rolling_apply in the for-loop:

grp['Result'] = pd.rolling_apply(grp['Value'], 60, between_1_3_perc)

answered Mar 18, 2013 at 21:55

herrfz

4,9044 gold badges29 silver badges37 bronze badges

Collectives™ on Stack Overflow

Customizing rolling_apply function in Python pandas

Setup

Rolling apply question

2 Answers 2

5 Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

Setup

Rolling apply question

2 Answers 2

5 Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related