Chunking
Why do we need chunks?
NannyML monitors ML models in production by doing data drift detection and performance estimation or monitoring. This functionality relies on aggregate metrics that are evaluated on samples of production data. These samples are called chunks.
All the results generated are calculated and presented per chunk i.e. a chunk is a single data point on the monitoring results. You can refer to Data Drift guide or Performance Estimation guide to review example results.
Walkthrough on creating chunks
To allow for flexibility there are many ways to create chunks. The examples below will show how different kinds of chunks can be created. The examples will be run based on the performance estimation flow on the synthetic binary classification dataset provided by NannyML. First we set up this dataset.
>>> import nannyml as nml
>>> reference, analysis, _ = nml.load_synthetic_car_loan_dataset()
Time-based chunking
Time-based chunking creates chunks based on time intervals. One chunk can contain all the observations
from a single hour, day, week, month etc. In most cases, such chunks will vary in the number of observations they
contain. Specify the chunk_period
argument to get appropriate split. The example below chunks data quarterly.
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
... chunk_period="Q", # here we define the chunk period.
>>> )
>>> cbpe.fit(reference)
>>> est_perf = cbpe.estimate(analysis)
>>> est_perf.data.iloc[:3, :6]
chunk
key
|
chunk_index
|
start_index
|
end_index
|
start_date
|
end_date
|
|
---|---|---|---|---|---|---|
0 |
2018Q1 |
0 |
0 |
14863 |
2018-01-01 00:00:00 |
2018-03-31 23:59:59.999999999 |
1 |
2018Q2 |
1 |
14864 |
29892 |
2018-04-01 00:00:00 |
2018-06-30 23:59:59.999999999 |
2 |
2018Q3 |
2 |
29893 |
45086 |
2018-07-01 00:00:00 |
2018-09-30 23:59:59.999999999 |
Note
Notice that each calendar quarter was taken into account, even if it was not fully covered with records. This means some chunks contain fewer observations (usually the last and the first). See the first row above - Q3 is July-September, but the first record in the data is from the last day of August. The first chunk has ~1200 of observations while the 2nd and 3rd contain above 3000. This can cause some chunks to be less reliably estimated or calculated.
Possible time offsets are listed in the table below:
Alias |
Description |
---|---|
S |
second |
T, min |
minute |
H |
hour |
D |
day |
W |
week |
M |
month |
Q |
quarter |
A, y |
year |
Size-based chunking
Chunks can be of fixed size, i.e. each chunk contains the same number of observations. Set this up by specifying the
chunk_size
parameter.
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
... chunk_size=3500, # here we define the chunk size.
>>> )
>>> cbpe.fit(reference)
>>> est_perf = cbpe.estimate(analysis)
>>> est_perf.data.iloc[:3, :6]
chunk
key
|
chunk_index
|
start_index
|
end_index
|
start_date
|
end_date
|
|
---|---|---|---|---|---|---|
0 |
[0:3499] |
0 |
0 |
3499 |
2018-01-01 00:00:00 |
2018-01-22 04:28:28.848000 |
1 |
[3500:6999] |
1 |
3500 |
6999 |
2018-01-22 04:37:12 |
2018-02-12 09:05:40.848000 |
2 |
[7000:10499] |
2 |
7000 |
10499 |
2018-02-12 09:14:24 |
2018-03-05 13:42:52.848000 |
Note
If the number of observations is not divisible by the chunk_size
required,
by default, the leftover observations will be appended to the last complete Chunk (overfilling it).
Notice that on the last chunk the difference between the start_index
and end_index
is greater than the chunk_size
defined.
Check the custom chunks section if you want to change the default behaviour.
>>> est_perf.data.iloc[-2:,:6]
chunk
key
|
chunk_index
|
start_index
|
end_index
|
start_date
|
end_date
|
|
---|---|---|---|---|---|---|
26 |
[42000:45499] |
12 |
42000 |
45499 |
2019-07-12 01:26:24 |
2019-08-02 05:54:52.848000 |
27 |
[45500:49999] |
13 |
45500 |
49999 |
2019-08-02 06:03:36 |
2019-08-29 11:51:16.848000 |
>>> last = est_perf.data.iloc[-1].loc['chunk']
>>> print(last.loc['end_index'] - last.loc['start_index'])
4499
Number-based chunking
The total number of chunks can be set by the chunk_number
parameter:
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
... chunk_number=9, # here we define the chunk number
>>> )
>>> cbpe.fit(reference)
>>> est_perf = cbpe.estimate(analysis)
>>> len(est_perf.filter(period='reference'))
9
Note
Chunks created this way will be equal in size.
If the number of observations is not divisible by the chunk_number
required, by default,
the leftover observations will be appended to the last complete Chunk (overfilling it).
Notice that on the last chunk the difference between the start_index and end_index is greater than the chunk_size defined.
Check the custom chunks section if you want to change the default behaviour.
>>> est_perf.filter(period='reference').data.iloc[-2:, :6]
chunk
key
|
chunk_index
|
start_index
|
end_index
|
start_date
|
end_date
|
|
---|---|---|---|---|---|---|
7 |
[38885:44439] |
7 |
38885 |
44439 |
2018-08-24 10:46:05.520000 |
2018-09-27 01:52:31.728000 |
8 |
[44440:49999] |
8 |
44440 |
49999 |
2018-09-27 02:01:14.880000 |
2018-10-30 17:51:16.848000 |
>>> first = est_perf.data.iloc[1].loc['chunk']
>>> last = est_perf.data.iloc[-1].loc['chunk']
>>> print('first chunk len:', first.loc['end_index'] - first.loc['start_index'])
>>> print('last chunk len:', last.loc['end_index'] - last.loc['start_index'])
first chunk len: 5554
last chunk len: 5559
Warning
The same splitting rule is always applied to the dataset used for fitting (reference
) and the dataset of
interest (in the presented case - analysis
).
Unless these two datasets are of the same size, the chunk sizes
can be considerably different. E.g. if the reference
dataset has 10 000 observations and the analysis
dataset has 80 000, and chunking is number-based, chunks in reference
will be much smaller than in
analysis
.
Additionally, if the data drift or performance estimation is calculated on
combined reference
and analysis
the results presented for reference
will be calculated on different
chunks than they were fitted.
Automatic chunking
The default chunking method is count-based, with the desired count set to 10. This is used if a chunking method isn’t specified.
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
>>> )
>>> cbpe.fit(reference)
>>> est_perf = cbpe.estimate(analysis)
>>> print(len(est_perf.filter(period='reference')))
10
Customize chunk behavior
A custom chunker
instance can be provided to change the default way of handling incomplete chunks,
or to handle a custom way of chunking the dataset.
For example, SizeBasedChunker
can be used to drop
the leftover observations to have fixed sized chunks.
>>> from nannyml.chunk import SizeBasedChunker, CountBasedChunker
>>> # The reference dataset contains 50000 records
>>> print(f"Size of reference data: {reference.shape[0]}")
>>> # We can use the 'drop' strategy to handle incomplete chunks
>>> chunker = SizeBasedChunker(chunk_size=3500 , incomplete='drop')
>>> last = chunker.split(reference)[-1]
>>> print(f"The last index: {last.end_index}")
>>> print(f"Last chunk size: {len(last)}")
Size of reference data: 50000
The last index: 48999
Last chunk size: 3500
You could also chunk your data into a fixed number of chunks, choosing to append any leftover observations to the last chunk.
>>> # The reference dataset contains 50000 records
>>> print(f"Size of reference data: {reference.shape[0]}")
>>> # We can use a different chunker with another 'incomplete' strategy
>>> chunker_count_drop = CountBasedChunker(chunk_number=9, incomplete='append')
>>> last = chunker_count_drop.split(reference)[-1]
>>> print(f"The last index: {last.end_index}")
>>> print(f"Last chunk size: {len(last)}")
Size of reference data: 50000
The last index: 49999
Last chunk size: 5560
You can then provide your custom chunker to the appropriate calculator or estimator.
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
... chunker=chunker_count_drop
>>> ).fit(reference_data=reference)
Chunks on plots with results
Finally, once the chunking method is selected, the full performance estimation can be run.
Each point on the plot represents a single chunk, with the y-axis showing the performance. They are aligned on the x axis with the date at the end of the chunk, not the date in the middle of the chunk. Plots are interactive - hovering over the point will display the precise information about the period, to help prevent any confusion.
>>> cbpe = nml.CBPE(
... y_pred_proba='y_pred_proba',
... y_pred='y_pred',
... y_true='repaid',
... timestamp_column_name='timestamp',
... metrics=['roc_auc'],
... problem_type='classification_binary',
... chunk_size=5_000
>>> ).fit(reference_data=reference)
>>> est_perf = cbpe.estimate(analysis)
>>> figure = est_perf.plot(kind='performance')
>>> figure.show()