Monitoring Realized Performance for Binary Classification

Note

The following example uses timestamps. These are optional but have an impact on the way data is chunked and results are plotted. You can read more about them in the data requirements.

Just The Code

>>> import nannyml as nml
>>> from IPython.display import display

>>> reference_df = nml.load_synthetic_binary_classification_dataset()[0]
>>> analysis_df = nml.load_synthetic_binary_classification_dataset()[1]
>>> analysis_target_df = nml.load_synthetic_binary_classification_dataset()[2]
>>> analysis_df = analysis_df.merge(analysis_target_df, on='identifier')

>>> display(reference_df.head(3))

>>> calc = nml.PerformanceCalculator(
...     y_pred_proba='y_pred_proba',
...     y_pred='y_pred',
...     y_true='work_home_actual',
...     timestamp_column_name='timestamp',
...     problem_type='classification_binary',
...     metrics=['roc_auc', 'f1', 'precision', 'recall', 'specificity', 'accuracy'],
...     chunk_size=5000)

>>> calc.fit(reference_df)

>>> results = calc.calculate(analysis_df)
>>> display(results.filter(period='analysis').to_df())

>>> display(results.filter(period='reference').to_df())

>>> figure = results.plot()
>>> figure.show()

Walkthrough

For simplicity this guide is based on a synthetic dataset included in the library, where the monitored model predicts whether an employee will work from home. You can read more about this synthetic dataset.

In order to monitor a model, NannyML needs to learn about it from a reference dataset. Then it can monitor the data that is subject to actual analysis, provided as the analysis dataset. You can read more about this in our section on data periods.

The analysis_targets dataframe contains the target results of the analysis period. This is kept separate in the synthetic data because it is not used during performance estimation.. But it is required to calculate performance, so the first thing we need to in this case is set up the right data in the right dataframes. The analysis target values are joined on the analysis frame by the identifier column.

>>> import nannyml as nml
>>> from IPython.display import display

>>> reference_df = nml.load_synthetic_binary_classification_dataset()[0]
>>> analysis_df = nml.load_synthetic_binary_classification_dataset()[1]
>>> analysis_target_df = nml.load_synthetic_binary_classification_dataset()[2]
>>> analysis_df = analysis_df.merge(analysis_target_df, on='identifier')

>>> display(reference_df.head(3))

	distance_from_office	salary_range	gas_price_per_litre	public_transportation_cost	wfh_prev_workday	workday	tenure	identifier	work_home_actual	timestamp	y_pred_proba	period	y_pred
0	5.96225	40K - 60K €	2.11948	8.56806	False	Friday	0.212653	0	1	2014-05-09 22:27:20	0.99	reference	1
1	0.535872	40K - 60K €	2.3572	5.42538	True	Tuesday	4.92755	1	0	2014-05-09 22:59:32	0.07	reference	0
2	1.96952	40K - 60K €	2.36685	8.24716	False	Monday	0.520817	2	1	2014-05-09 23:48:25	1	reference	1

Next a PerformanceCalculator is created using a list of metrics to calculate (or just one metric), the data columns required for these metrics, and an optional chunking specification.

The list of metrics specifies which performance metrics of the monitored model will be calculated. The following metrics are currently supported:

• roc_auc - one-vs-the-rest, macro-averaged

• f1 - macro-averaged

• precision - macro-averaged

• recall - macro-averaged

• specificity - macro-averaged

• accuracy

For more information on metrics, check the metrics module.

>>> calc = nml.PerformanceCalculator(
...     y_pred_proba='y_pred_proba',
...     y_pred='y_pred',
...     y_true='work_home_actual',
...     timestamp_column_name='timestamp',
...     problem_type='classification_binary',
...     metrics=['roc_auc', 'f1', 'precision', 'recall', 'specificity', 'accuracy'],
...     chunk_size=5000)

>>> calc.fit(reference_df)

The new PerformanceCalculator is fitted using the fit() method on the reference data.

The fitted PerformanceCalculator can then be used to calculate realized performance metrics on all data which has target values available with the calculate() method. NannyML can output a dataframe that contains all the results of the analysis data.

>>> results = calc.calculate(analysis_df)
>>> display(results.filter(period='analysis').to_df())

	(‘chunk’, ‘key’)	(‘chunk’, ‘chunk_index’)	(‘chunk’, ‘start_index’)	(‘chunk’, ‘end_index’)	(‘chunk’, ‘start_date’)	(‘chunk’, ‘end_date’)	(‘chunk’, ‘period’)	(‘chunk’, ‘targets_missing_rate’)	(‘roc_auc’, ‘sampling_error’)	(‘roc_auc’, ‘value’)	(‘roc_auc’, ‘upper_threshold’)	(‘roc_auc’, ‘lower_threshold’)	(‘roc_auc’, ‘alert’)	(‘f1’, ‘sampling_error’)	(‘f1’, ‘value’)	(‘f1’, ‘upper_threshold’)	(‘f1’, ‘lower_threshold’)	(‘f1’, ‘alert’)	(‘precision’, ‘sampling_error’)	(‘precision’, ‘value’)	(‘precision’, ‘upper_threshold’)	(‘precision’, ‘lower_threshold’)	(‘precision’, ‘alert’)	(‘recall’, ‘sampling_error’)	(‘recall’, ‘value’)	(‘recall’, ‘upper_threshold’)	(‘recall’, ‘lower_threshold’)	(‘recall’, ‘alert’)	(‘specificity’, ‘sampling_error’)	(‘specificity’, ‘value’)	(‘specificity’, ‘upper_threshold’)	(‘specificity’, ‘lower_threshold’)	(‘specificity’, ‘alert’)	(‘accuracy’, ‘sampling_error’)	(‘accuracy’, ‘value’)	(‘accuracy’, ‘upper_threshold’)	(‘accuracy’, ‘lower_threshold’)	(‘accuracy’, ‘alert’)
0	[0:4999]	0	0	4999	2017-08-31 04:20:00	2018-01-02 00:45:44	analysis	0	0.00181072	0.970962	0.97866	0.963317	False	0.00610429	0.949549	0.961094	0.935047	False	0.00461594	0.942139	0.961131	0.924741	False	0.00422251	0.957077	0.965726	0.940831	False	0.00465744	0.937034	0.960113	0.924741	False	0.0031445	0.9474	0.960601	0.935079	False
1	[5000:9999]	1	5000	9999	2018-01-02 01:13:11	2018-05-01 13:10:10	analysis	0	0.00181072	0.970248	0.97866	0.963317	False	0.00610429	0.946686	0.961094	0.935047	False	0.00461594	0.943434	0.961131	0.924741	False	0.00422251	0.949959	0.965726	0.940831	False	0.00465744	0.944925	0.960113	0.924741	False	0.0031445	0.9474	0.960601	0.935079	False
2	[10000:14999]	2	10000	14999	2018-05-01 14:25:25	2018-09-01 15:40:40	analysis	0	0.00181072	0.976282	0.97866	0.963317	False	0.00610429	0.950459	0.961094	0.935047	False	0.00461594	0.941438	0.961131	0.924741	False	0.00422251	0.959654	0.965726	0.940831	False	0.00465744	0.943602	0.960113	0.924741	False	0.0031445	0.9514	0.960601	0.935079	False
3	[15000:19999]	3	15000	19999	2018-09-01 16:19:07	2018-12-31 10:11:21	analysis	0	0.00181072	0.967721	0.97866	0.963317	False	0.00610429	0.945968	0.961094	0.935047	False	0.00461594	0.946731	0.961131	0.924741	False	0.00422251	0.945205	0.965726	0.940831	False	0.00465744	0.947577	0.960113	0.924741	False	0.0031445	0.9464	0.960601	0.935079	False
4	[20000:24999]	4	20000	24999	2018-12-31 10:38:45	2019-04-30 11:01:30	analysis	0	0.00181072	0.969886	0.97866	0.963317	False	0.00610429	0.944136	0.961094	0.935047	False	0.00461594	0.940039	0.961131	0.924741	False	0.00422251	0.948269	0.965726	0.940831	False	0.00465744	0.938882	0.960113	0.924741	False	0.0031445	0.9436	0.960601	0.935079	False
5	[25000:29999]	5	25000	29999	2019-04-30 11:02:00	2019-09-01 00:24:27	analysis	0	0.00181072	0.96005	0.97866	0.963317	True	0.00610429	0.915794	0.961094	0.935047	True	0.00461594	0.88822	0.961131	0.924741	True	0.00422251	0.945134	0.965726	0.940831	False	0.00465744	0.881342	0.960113	0.924741	True	0.0031445	0.9132	0.960601	0.935079	True
6	[30000:34999]	6	30000	34999	2019-09-01 00:28:54	2019-12-31 09:09:12	analysis	0	0.00181072	0.95853	0.97866	0.963317	True	0.00610429	0.920015	0.961094	0.935047	True	0.00461594	0.898152	0.961131	0.924741	True	0.00422251	0.94297	0.965726	0.940831	False	0.00465744	0.890909	0.960113	0.924741	True	0.0031445	0.9172	0.960601	0.935079	True
7	[35000:39999]	7	35000	39999	2019-12-31 10:07:15	2020-04-30 11:46:53	analysis	0	0.00181072	0.959041	0.97866	0.963317	True	0.00610429	0.915063	0.961094	0.935047	True	0.00461594	0.890992	0.961131	0.924741	True	0.00422251	0.940471	0.965726	0.940831	True	0.00465744	0.884662	0.960113	0.924741	True	0.0031445	0.9126	0.960601	0.935079	True
8	[40000:44999]	8	40000	44999	2020-04-30 12:04:32	2020-09-01 02:46:02	analysis	0	0.00181072	0.963094	0.97866	0.963317	True	0.00610429	0.922835	0.961094	0.935047	True	0.00461594	0.902232	0.961131	0.924741	True	0.00422251	0.9444	0.965726	0.940831	False	0.00465744	0.899126	0.960113	0.924741	True	0.0031445	0.9216	0.960601	0.935079	True
9	[45000:49999]	9	45000	49999	2020-09-01 02:46:13	2021-01-01 04:29:32	analysis	0	0.00181072	0.957556	0.97866	0.963317	True	0.00610429	0.914221	0.961094	0.935047	True	0.00461594	0.886848	0.961131	0.924741	True	0.00422251	0.943337	0.965726	0.940831	False	0.00465744	0.873822	0.960113	0.924741	True	0.0031445	0.9094	0.960601	0.935079	True

There results from the reference data are also available.

>>> display(results.filter(period='reference').to_df())

	(‘chunk’, ‘key’)	(‘chunk’, ‘chunk_index’)	(‘chunk’, ‘start_index’)	(‘chunk’, ‘end_index’)	(‘chunk’, ‘start_date’)	(‘chunk’, ‘end_date’)	(‘chunk’, ‘period’)	(‘chunk’, ‘targets_missing_rate’)	(‘roc_auc’, ‘sampling_error’)	(‘roc_auc’, ‘value’)	(‘roc_auc’, ‘upper_threshold’)	(‘roc_auc’, ‘lower_threshold’)	(‘roc_auc’, ‘alert’)	(‘f1’, ‘sampling_error’)	(‘f1’, ‘value’)	(‘f1’, ‘upper_threshold’)	(‘f1’, ‘lower_threshold’)	(‘f1’, ‘alert’)	(‘precision’, ‘sampling_error’)	(‘precision’, ‘value’)	(‘precision’, ‘upper_threshold’)	(‘precision’, ‘lower_threshold’)	(‘precision’, ‘alert’)	(‘recall’, ‘sampling_error’)	(‘recall’, ‘value’)	(‘recall’, ‘upper_threshold’)	(‘recall’, ‘lower_threshold’)	(‘recall’, ‘alert’)	(‘specificity’, ‘sampling_error’)	(‘specificity’, ‘value’)	(‘specificity’, ‘upper_threshold’)	(‘specificity’, ‘lower_threshold’)	(‘specificity’, ‘alert’)	(‘accuracy’, ‘sampling_error’)	(‘accuracy’, ‘value’)	(‘accuracy’, ‘upper_threshold’)	(‘accuracy’, ‘lower_threshold’)	(‘accuracy’, ‘alert’)
0	[0:4999]	0	0	4999	2014-05-09 22:27:20	2014-09-09 08:18:27	reference	0	0.00181072	0.976253	0.97866	0.963317	False	0.00610429	0.953803	0.961094	0.935047	False	0.00461594	0.951308	0.961131	0.924741	False	0.00422251	0.956311	0.965726	0.940831	False	0.00465744	0.952136	0.960113	0.924741	False	0.0031445	0.9542	0.960601	0.935079	False
1	[5000:9999]	1	5000	9999	2014-09-09 09:13:35	2015-01-09 00:02:51	reference	0	0.00181072	0.969045	0.97866	0.963317	False	0.00610429	0.940963	0.961094	0.935047	False	0.00461594	0.934748	0.961131	0.924741	False	0.00422251	0.947262	0.965726	0.940831	False	0.00465744	0.9357	0.960113	0.924741	False	0.0031445	0.9414	0.960601	0.935079	False
2	[10000:14999]	2	10000	14999	2015-01-09 00:04:43	2015-05-09 15:54:26	reference	0	0.00181072	0.971742	0.97866	0.963317	False	0.00610429	0.954483	0.961094	0.935047	False	0.00461594	0.949804	0.961131	0.924741	False	0.00422251	0.959208	0.965726	0.940831	False	0.00465744	0.948283	0.960113	0.924741	False	0.0031445	0.9538	0.960601	0.935079	False
3	[15000:19999]	3	15000	19999	2015-05-09 16:02:08	2015-09-07 07:14:37	reference	0	0.00181072	0.971642	0.97866	0.963317	False	0.00610429	0.946237	0.961094	0.935047	False	0.00461594	0.941363	0.961131	0.924741	False	0.00422251	0.951161	0.965726	0.940831	False	0.00465744	0.940847	0.960113	0.924741	False	0.0031445	0.946	0.960601	0.935079	False
4	[20000:24999]	4	20000	24999	2015-09-07 07:27:47	2016-01-08 16:02:05	reference	0	0.00181072	0.969085	0.97866	0.963317	False	0.00610429	0.944324	0.961094	0.935047	False	0.00461594	0.942285	0.961131	0.924741	False	0.00422251	0.946372	0.965726	0.940831	False	0.00465744	0.940341	0.960113	0.924741	False	0.0031445	0.9434	0.960601	0.935079	False
5	[25000:29999]	5	25000	29999	2016-01-08 17:22:00	2016-05-09 11:09:39	reference	0	0.00181072	0.967364	0.97866	0.963317	False	0.00610429	0.945286	0.961094	0.935047	False	0.00461594	0.937525	0.961131	0.924741	False	0.00422251	0.953176	0.965726	0.940831	False	0.00465744	0.938679	0.960113	0.924741	False	0.0031445	0.9458	0.960601	0.935079	False
6	[30000:34999]	6	30000	34999	2016-05-09 11:19:36	2016-09-04 03:30:35	reference	0	0.00181072	0.968692	0.97866	0.963317	False	0.00610429	0.94885	0.961094	0.935047	False	0.00461594	0.939168	0.961131	0.924741	False	0.00422251	0.958734	0.965726	0.940831	False	0.00465744	0.938099	0.960113	0.924741	False	0.0031445	0.9484	0.960601	0.935079	False
7	[35000:39999]	7	35000	39999	2016-09-04 04:09:35	2017-01-03 18:48:21	reference	0	0.00181072	0.970205	0.97866	0.963317	False	0.00610429	0.948262	0.961094	0.935047	False	0.00461594	0.940831	0.961131	0.924741	False	0.00422251	0.955812	0.965726	0.940831	False	0.00465744	0.939309	0.960113	0.924741	False	0.0031445	0.9476	0.960601	0.935079	False
8	[40000:44999]	8	40000	44999	2017-01-03 19:00:51	2017-05-03 02:34:24	reference	0	0.00181072	0.974096	0.97866	0.963317	False	0.00610429	0.953456	0.961094	0.935047	False	0.00461594	0.953645	0.961131	0.924741	False	0.00422251	0.953267	0.965726	0.940831	False	0.00465744	0.952727	0.960113	0.924741	False	0.0031445	0.953	0.960601	0.935079	False
9	[45000:49999]	9	45000	49999	2017-05-03 02:49:38	2017-08-31 03:10:29	reference	0	0.00181072	0.971757	0.97866	0.963317	False	0.00610429	0.945042	0.961094	0.935047	False	0.00461594	0.938687	0.961131	0.924741	False	0.00422251	0.951484	0.965726	0.940831	False	0.00465744	0.938148	0.960113	0.924741	False	0.0031445	0.9448	0.960601	0.935079	False

Apart from chunking and chunk and period-related columns, the results data have a set of columns for each calculated metric. When taking roc_auc as an example:

• targets_missing_rate - The fraction of missing target data.

• <metric> - The value of the metric for a specific chunk.

• <metric>_lower_threshold> and <metric>_upper_threshold> - Lower and upper thresholds for performance metric. Crossing them will raise an alert that there is a significant metric change. The thresholds are calculated based on the realized performance of chunks in the reference period. The thresholds are 3 standard deviations away from the mean performance calculated on reference chunks. They are calculated during fit phase.

• <metric>_alert - A flag indicating potentially significant performance change. True if realized performance crosses upper or lower threshold.

• <metric>_sampling_error - Estimated Sampling Error for the relevant metric.

The results can be plotted for visual inspection.

>>> figure = results.plot()
>>> figure.show()

Insights

After reviewing the performance calculation results, we should be able to clearly see how the model is performing against the targets, according to whatever metrics we wish to track.

What Next

If we decide further investigation is needed, the Data Drift functionality can help us to see what feature changes may be contributing to any performance changes.

It is also wise to check whether the model’s performance is satisfactory according to business requirements. This is an ad-hoc investigation that is not covered by NannyML.