Column mapping
How to use column mapping in Evidently.
TL;DR: Evidently expects a certain dataset structure and input column names. You can specify any differences by creating a ColumnMapping object. It works the same way for test suites and reports.
Column mapping helps correctly process the input data.
If the
column_mapping is not specified or set as None, Evidently will use the default mapping strategy.Column types:
- All columns with numeric types (np.number) will be treated as Numerical.
- All columns with DateTime format (np.datetime64) will be treated as DateTime.
- All other columns will be treated as Categorical.
Dataset structure:
- The column named âidâ will be treated as an ID column.
- The column named âdatetimeâ will be treated as a DateTime column.
- The column named âtargetâ will be treated as a target function.
- The column named âpredictionâ will be treated as a model prediction.
ID, DateTime, target, and prediction are utility columns. If provided, the âdatetimeâ column will be used as an index for some plots. The âIDâ column will be excluded from drift analysis.
Data structure requirements depend on the type of analysis. Here are example requirements:
Report/Test Suite | Feature columns | Prediction column | Target column | ID column | Datetime column |
|---|---|---|---|---|---|
Data Quality | Required | Optional | Optional | Optional | Optional |
Data Drift | Required | Optional | Optional | Optional | Optional |
Target Drift | Optional | Target and/or prediction required | Target and/or prediction required | Optional | Optional |
Classification Performance | Optional | Required | Required | Optional | Optional |
Regression Performance | Optional | Required | Required | Optional | Optional |
We recommend specifying column mapping manually. Evidently applies different heuristics and rules to map the input data automatically. To avoid errors, it is always best to set column mapping manually. For example, numerical columns with only 3 different values in the reference data might be incorrectly parsed as categorical features.
You can create a
ColumnMapping object to map your column names and feature types.To specify the column names for target, prediction, ID or DateTime:
from evidently import ColumnMapping
â
column_mapping = ColumnMapping()
â
column_mapping.target = 'y' #'y' is the name of the column with the target function
column_mapping.prediction = 'pred' #'pred' is the name of the column(s) with model predictions
column_mapping.id = None #there is no ID column in the dataset
column_mapping.datetime = 'date' #'date' is the name of the column with datetime
To split the features into numerical and categorical types:
column_mapping.numerical_features = ['temp', 'atemp', 'humidity'] #list of numerical features
column_mapping.categorical_features = ['season', 'holiday'] #list of categorical features
Why map them: the column types affect some of the tests, metrics and visualizations. For example, the drift algorithm selects a statistical test based on the column type and ignores DateTime features. Some of the data quality visualizations are different for specific feature types. Some of the tests (e.g. on value ranges) only considers numeral columns, etc.
There are additional mapping options that apply to specific test suites and reports.
You might have temporal features in your dataset. For example, âdate of the last contact.â
To map them, define:
column_mapping = ColumnMapping()
column_mapping.datetime_features = ['last_call_date', 'join_date'] #list of DateTime features
Default: Evidently treats columns with DateTime format (np.datetime64) as DateTime features.
Note: do not confuse DateTime features with the DateTime column, which is used as the x-axis in some plots. You will typically use the DateTime column as a prediction timestamp.
Why map them: if you specify DateTime features, they will be ignored in data drift calculation. Evidently will also calculate appropriate statistics and generate different visualizations for DateTime features in the data quality report.
In many cases, it is important to differentiate between continuous and discrete targets. This applies to multiple reports and tests, including Data Quality and Target Drift.
To define it explicitly, specify the task parameter:
column_mapping = ColumnMapping()
column_mapping.target = 'y'
column_mapping.task = 'regression'
It accepts two values: 'regression' and 'classification'.
Default: If you don't specify the task, Evidently will use a simple strategy: if the target has a numeric type and the number of unique values > 5: task == âregression.â In all other cases, the task == âclassificationâ.
Why map it: If you have a multi-class problem where classes are encoded as numbers, it might look the same way as a regression problem. Thus it is best to explicitly specify it. It will affect how the target (prediction) is visualized and help pick the correct statistical tests for the target (prediction) drift detection. It will also affect the calculation of statistics and tests that differ for numerical and categorical data types.
To evaluate the classification performance, you need both true labels and prediction. Depending on the classification type (e.g., binary, multi-class, probabilistic), you have different options of how to pass the predictions.
Target: encoded labels, Preds: encoded labels + Optional[target_names].
target | prediction |
|---|---|
1 | 1 |
0 | 2 |
⊠| ⊠|
2 | 2 |
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'prediction'
column_mapping.target_names = ['Setosa', 'Versicolour', 'Virginica']
If you pass the target names, they will appear on the visualizations.
Target: labels, Preds: labels.
target | prediction |
|---|---|
âVersicolourâ | âVersicolourâ |
âSetosaâ | âVirginicaâ |
⊠| ⊠|
âVirginicaâ | âVirginicaâ |
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'prediction'
Target: labels, Preds: columns named after labels.
target | âVersicolourâ | âSetosaâ | âVirginicaâ |
|---|---|---|---|
âSetosaâ | 0.98 | 0.01 | 0.01 |
âVirginicaâ | 0.5 | 0.2 | 0.3 |
⊠| ⊠| â | â |
âVirginicaâ | 0.2 | 0.7 | 0.1 |
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = ['Setosa', 'Versicolour', 'Virginica']
Naming the columns after the lables is a requirement. You cannot pass a custom list.
Target: encoded labels, Preds: encoded labels + pos_label + Optional[target_names]
target | prediction |
|---|---|
1 | 1 |
0 | 1 |
⊠| ⊠|
1 | 0 |
By default, Evidently expects the positive class to be labeled as â1â. If you have a different label, specify it explicitly.
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'prediction'
column_mapping.target_names = ['churn', 'not_churn']
pos_label = 0
If you pass the target names, they will appear on the visualizations.
Target: labels, Preds: labels + pos_label
target | prediction |
|---|---|
âchurnâ | âchurnâ |
ânot_churnâ | âchurnâ |
⊠| ⊠|
âchurnâ | ânot_churnâ |
Passing the name of the positive class is a requirement in this case.
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'prediction'
pos_label = 'churn'
Target: labels, Preds: columns named after labels + pos_label
target | âchurnâ | ânot_churnâ |
|---|---|---|
âchurnâ | 0.9 | 0.1 |
âchurnâ | 0.7 | 0.3 |
⊠| ⊠| â |
ânot_churnâ | 0.5 | 0.5 |
Passing the name of the positive class is a requirement in this case.
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = ['churn', 'not_churn']
pos_label = 'churn'
Target: labels, Preds: a column named like one of the labels + pos_label
target | ânot_churnâ |
|---|---|
âchurnâ | 0.5 |
ânot_churnâ | 0.1 |
⊠| ⊠|
âchurnâ | 0.9 |
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'not_churn'
pos_label = 'churn'
Both naming the column after one of the labels and passing the name of the positive class are requirements.
Target: encoded labels, Preds: one column with any name + pos_label
target | prediction |
|---|---|
1 | 0.5 |
1 | 0.1 |
⊠| ⊠|
0 | 0.9 |
column_mapping = ColumnMapping()
â
column_mapping.target = 'target'
column_mapping.prediction = 'prediction'
pos_label = 1
column_mapping.target_names = ['churn', 'not_churn']
If you pass the target names, they will appear on the visualizations.
Last modified 21d ago