Pandas Groupby Operation For Condition Based Feature Creation

I think we could apply here sort of categorical logic. Here’s what I mean.

Let’s say ['type3','type1','type2','both_types'] are verbal representatives of codes [0, 1, 2, 3]. Why this order? As I can see, type3 is sort of a default value. type1 and type2 are equal by nature, but we have somehow to differ them. So let’s follow their names and say that type1 is 1 and type2 is 2. And both_types has index 3 as a sum of previos two codes. Now we can separate identifying if a record can be of type 1 or 2, and get their sum as a final output. If a record is neither type 1 nor type 2 kind of error, the sum will result in 0. If there’s only one of them, the sum will keep it as is. If a record can be of both types, the sum will show 3.

Let’s see how it looks in code:

err = custom_df['err_type'] == 'UNKNOWN'
pli = custom_df['pli_missing'] == 'yes'
grouper = custom_df['ir_id']

# with transform get the same dimension as custom_df
type1 = (err & pli).groupby(grouper).transform(any)
type2 = (err & ~pli).groupby(grouper).transform(any)

codes = type1 + 2*type2
categories = ['type3','type1','type2','both_types']
custom_df['err_flag_new'] = pd.Categorical.from_codes(codes, categories)

Here’s what I’ve got in the end:

update

We can look at it this way. Suppose there’s a system with some number of independent states. A system can be described as having any combination of them. It can be described mathematically as a binary code. For each independant state we assing a unique place in this code, where values 1 or 0 are interpreted as a logical answer whether the system has a corresponding state. Binary means a sum like this one:

state[0]*2^0 + state[1]*2^1 + state[2]*2^2 + state[3]*2^3 + ...

In our case we have only 2 independent states type1 and type2. Two others are their combinations: type3 means neather of them, and both_types tells for itself. So we have only first two terms of the sum above, where state[0] is a logical value for type1 and state[1] is a logical value for type2. That’s why I used codes = type1 + 2*type2 which is equal to codes = type1 * 2**0 + type2 * 2**1

As for the order in ['type3','type1','type2','both_types'], in this list indexes of values resemble the corresponding codes, ie. each type here has the index equal to their binary code. The binary code for type3 in this model is 0b00 which is zero, for type1 it is 0b01 which is one, for type2 – 0b10 which is 2 and for both_types it’s 0b11 which is equal to 3. These codes are atomatically assigned when creating Categorical sequence from codes, ie. pandas is using codes as indexes to get corresponding values from the list and place them instead of the codes.

See also Enum.IntFlag as an abstract realization of this idea, and Flags in the regular expression module as an example of how it can be used.