preprocessing.RepeatingBasisFunction encodes different labels as the same[BUG]

[yug030]

Hi,

I'm working on a task to encode timestamps (more specifically, weekdays) and I noticed RepeatingBasisFunction does not work correctly with weekday=0 and weekday=6: the outputs for both are the same.

Here's the code snippet to reproduce the bug:

import pandas as pd
from sklego.preprocessing import RepeatingBasisFunction

start_date = '2023-06-25 00:00:00'
end_date = '2023-07-03 23:59:59'
timestamps = pd.date_range(start=start_date, end=end_date, freq='1D')
df = pd.DataFrame({'Timestamps': timestamps})
df['wkday'] = df['Timestamps'].dt.weekday
print(df)
rbf = RepeatingBasisFunction(n_periods=4, column="wkday", input_range=(0,6), remainder="drop")
rbf.fit(df)
print(rbf.transform(df))

The output for me is:

  Timestamps  wkday
0 2023-06-25      6
1 2023-06-26      0
2 2023-06-27      1
3 2023-06-28      2
4 2023-06-29      3
5 2023-06-30      4
6 2023-07-01      5
7 2023-07-02      6
8 2023-07-03      0

array([[1.        , 0.36787944, 0.01831564, 0.36787944],
       [1.        , 0.36787944, 0.01831564, 0.36787944],
       [0.64118039, 0.89483932, 0.16901332, 0.06217652],
       [0.16901332, 0.89483932, 0.64118039, 0.06217652],
       [0.01831564, 0.36787944, 1.        , 0.36787944],
       [0.16901332, 0.06217652, 0.64118039, 0.89483932],
       [0.64118039, 0.06217652, 0.16901332, 0.89483932],
       [1.        , 0.36787944, 0.01831564, 0.36787944],
       [1.        , 0.36787944, 0.01831564, 0.36787944]])

---

Update:
If I add 1 to "wkday" column (so it's 1 to 7 now) and use rbf = RepeatingBasisFunction(n_periods=7, column="wkday", input_range=(0,7), remainder="drop") and call fit and transform, I will get a square matrix with 1.0 on the diagonal which looks correct, but it's different from the documentation.

[FBruzzesi]

Hi, thank you for reporting the issue and sorry for the late reply!

This seems to be unexpected/buggy behavior. My understanding is that some assumptions have been made in the implementation and it may be worth to adjust or clearly state them.

I will take the time to have a closer look at it.

[Alex-Cremers]

It's a bit confusing, but I think that this is expected behavior. Theinput_range argument is to be understood as an interval that loops on itself at the boundaries, so if you provide (0,6) you assume that 6 is the same as 0, leaving only distinct 6 integers in the range. In this case, you want to provide (0,7), which translates to the interval [0, 7) and does contain 7 distinct integers.

If you think in terms of continuous variables rather than integer, it possibly makes more sense.

[C3L0]

Hi @FBruzzesi @koaning!
I took a look at sklego/preprocessing/repeating_basis_function.py.

It seems that the cause of the issue #574 lies in lines 193-195 of _RepeatingBasisFunction:

if self.input_range is None:
    self.input_range = (X.min(), X.max())

Because normalized values 0.0 and 1.0 are defined as identical points on the repeating curve, setting input_range = (X.min(), X.max()) by default guarantees that X.min() and X.max() map to identical outputs. While RepeatingBasisFunction is mathematically correct for continuous periodic intervals $[a, b)$, this default (min, max) fallback causes boundary collisions when working with discrete categories like weekdays ($0 \dots 6$).

I'd like to work on this issue

I propose to :

• Clarify in the RepeatingBasisFunction docstrings that the upper bound of input_range is the wrap-around point.
• Update the docstring for input_range=None so users are explicitly warned that automatic (X.min(), X.max()) inference will collapse the boundary values onto each other.
• Add a unit test verifying that full-period ranges produce distinct feature encodings across discrete values.

Let me know if this direction works and I'll submit a PR!