Train Test Feature Drift#

This notebooks provides an overview for using and understanding feature drift check.


What is a feature drift?#

Drift is simply a change in the distribution of data over time, and it is also one of the top reasons why machine learning model’s performance degrades over time.

Feature drift is a data drift that occurs in a single feature in the dataset.

For more information on drift, please visit our drift guide.

How Deepchecks Detects Feature Drift#

This check detects feature drift by using univariate measures on each feature column separately. Another possible method for drift detection is by a domain classifier which is used in the Multivariate Drift check.

Generate data & model#

Let’s generate a mock dataset of 2 categorical and 2 numerical features

import numpy as np
import pandas as pd


train_data = np.concatenate([np.random.randn(1000,2), np.random.choice(a=['apple', 'orange', 'banana'], p=[0.5, 0.3, 0.2], size=(1000, 2))], axis=1)
test_data = np.concatenate([np.random.randn(1000,2), np.random.choice(a=['apple', 'orange', 'banana'], p=[0.5, 0.3, 0.2], size=(1000, 2))], axis=1)

df_train = pd.DataFrame(train_data, columns=['numeric_without_drift', 'numeric_with_drift', 'categorical_without_drift', 'categorical_with_drift'])
df_test = pd.DataFrame(test_data, columns=df_train.columns)

df_train = df_train.astype({'numeric_without_drift': 'float', 'numeric_with_drift': 'float'})
df_test = df_test.astype({'numeric_without_drift': 'float', 'numeric_with_drift': 'float'})
numeric_without_drift numeric_with_drift categorical_without_drift categorical_with_drift
0 0.496714 -0.138264 apple apple
1 0.647689 1.523030 apple apple
2 -0.234153 -0.234137 banana banana
3 1.579213 0.767435 apple banana
4 -0.469474 0.542560 orange apple

Insert drift into test:#

Now, we insert a synthetic drift into 2 columns in the dataset

df_test['numeric_with_drift'] = df_test['numeric_with_drift'].astype('float') + abs(np.random.randn(1000)) + np.arange(0, 1, 0.001) * 4
df_test['categorical_with_drift'] = np.random.choice(a=['apple', 'orange', 'banana', 'lemon'], p=[0.5, 0.25, 0.15, 0.1], size=(1000, 1))

Training a model#

Now, we are building a dummy model (the label is just a random numerical column). We preprocess our synthetic dataset so categorical features are being encoded with an OrdinalEncoder

from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OrdinalEncoder
from sklearn.tree import DecisionTreeClassifier

from deepchecks.tabular import Dataset
model = Pipeline([
    ('handle_cat', ColumnTransformer(
            ('num', 'passthrough',
             ['numeric_with_drift', 'numeric_without_drift']),
                 ('encode', OrdinalEncoder(handle_unknown='use_encoded_value', unknown_value=-1)),
             ['categorical_with_drift', 'categorical_without_drift'])
    ('model', DecisionTreeClassifier(random_state=0, max_depth=2))]
label = np.random.randint(0, 2, size=(df_train.shape[0],))
cat_features = ['categorical_without_drift', 'categorical_with_drift']
df_train['target'] = label
train_dataset = Dataset(df_train, label='target', cat_features=cat_features)[train_dataset.features], label)

label = np.random.randint(0, 2, size=(df_test.shape[0],))
df_test['target'] = label
test_dataset = Dataset(df_test, label='target', cat_features=cat_features)

Run the check#

Let’s run deepchecks’ feature drift check and see the results

from deepchecks.tabular.checks import TrainTestFeatureDrift

check = TrainTestFeatureDrift()
result =, test_dataset=test_dataset, model=model)
Train Test Feature Drift

Observe the check’s output#

As we see from the results, the check detects and returns the drift score per feature. As we expect, the features that were manually manipulated to contain a strong drift in them were detected.

In addition to the graphs, each check returns a value that can be controlled in order to define expectations on that value (for example, to define that the drift score for every feature must be below 0.05).

Let’s see the result value for our check

OrderedDict([('numeric_without_drift', {'Drift score': 0.019594833552359095, 'Method': "Earth Mover's Distance", 'Importance': 0.6911764705882353}), ('numeric_with_drift', {'Drift score': 0.3430867349314306, 'Method': "Earth Mover's Distance", 'Importance': 0.3088235294117647}), ('categorical_without_drift', {'Drift score': 0.005136700975462043, 'Method': "Cramer's V", 'Importance': 0.0}), ('categorical_with_drift', {'Drift score': 0.22862322289807285, 'Method': "Cramer's V", 'Importance': 0.0})])

Define a condition#

As we can see, we get the drift score for each feature in the dataset, along with the feature importance in respect to the model.

Now, we define a condition that enforce each feature’s drift score must be below 0.1. A condition is deepchecks’ way to enforce that results are OK, and we don’t have a problem in our data or model!

check_cond = check.add_condition_drift_score_less_than(max_allowed_categorical_score=0.2,
result =, test_dataset=test_dataset)
Train Test Feature Drift

As we see, our condition successfully detects and filters the problematic features that contains a drift!

Get an aggregated value#

Using the reduce_output function we can combine the drift values per feature and get a collective score that reflects the effect of the drift on the model, taking into account all the features. In scenarios where labels are unavailable (either temporarily of permanently) this value can be a good indicator of possible deterioration in the model’s performance.

We can define the type of aggregation we want to use via the aggregation_method parameter. The possible values are:

weighted: The default. Weighted mean of drift scores based on each feature’s feature importance. This method underlying logic is that drift in a feature with a higher feature importance will have a greater effect on the model’s performance.

mean: Simple mean of all the features drift scores.

none: No averaging. Return a dict with a drift score for each feature.

max: Maximum value of all the individual feature’s drift scores.

check = TrainTestFeatureDrift(aggregation_method='weighted')
result =, test_dataset=test_dataset, model=model)
{'Weighted Drift Score': 0.11949674427236648}

Total running time of the script: ( 0 minutes 1.057 seconds)

Gallery generated by Sphinx-Gallery