Data Drift
TL;DR: The report detects changes in feature distributions.
  • Performs a suitable statistical test for numerical and categorical features
  • Plots feature values and distributions for the two datasets.

Summary

The Data Drift report helps detect and explore changes in the input data.

Requirements

You will need two datasets. The reference dataset serves as a benchmark. We analyze the change by comparing the current production data to the reference data.
The dataset should include the features you want to evaluate for drift. The schema of both datasets should be identical.
  • In the case of pandas DataFrame,all column names should be string
  • All feature columns analyzed for drift should have the numerical type (np.number)
  • Categorical data can be encoded as numerical labels and specified in the column_mapping.
  • DateTime column is the only exception. If available, it can be used as the x-axis in the plots.
You can potentially choose any two datasets for comparison. But keep in mind that only the reference dataset will be used as a basis for comparison.

How it works

To estimate the data drift Evidently compares the distributions of each feature in the two datasets.
Evidently applies statistical tests to detect if the distribution has changed significantly. There is a default logic to choosing the appropriate statistical test based on:
  • feature type: categorical or numerical
  • the number of observations in the reference dataset
  • the number of unique values in the feature (n_unique)
For small data with <= 1000 observations in the reference dataset:
  • For numerical features (n_unique > 5): two-sample Kolmogorov-Smirnov test.
  • For categorical features or numerical features with n_unique <= 5: chi-squared test.
  • For binary categorical features (n_unique <= 2), we use the proportion difference test for independent samples based on Z-score.
All tests use a 0.95 confidence level by default.
For larger data with > 1000 observations in the reference dataset:
All tests use a threshold = 0.1 by default.
You can modify the drift detection logic by selecting a statistical test already available in the library, including PSI, K–L divergence, Jensen-Shannon distance, Wasserstein distance. See more details about available tests. You can also set a different confidence level or implement a custom test, by defining custom options.
To build up a better intuition for which tests are better in different kinds of use cases, visit our blog to read our in-depth guide to the tradeoffs when choosing the statistical test for data drift.

How it looks

The default report includes 4 components. All plots are interactive.

1. Data Drift Summary

The report returns the share of drifting features and an aggregate Dataset Drift result. For example:
Dataset Drift sets a rule on top of the results of the statistical tests for individual features. By default, Dataset Drift is detected if at least 50% of features drift at a 0.95 confidence level.
To set different Dataset Drift conditions, you can define custom options.

2. Data Drift Table

The table shows the drifting features first, sorting them by P-value. You can also choose to sort the rows by the feature name or type.

3. Data Drift by Feature

By clicking on each feature, you can explore the values mapped in a plot.
  • The dark green line is the mean, as seen in the reference dataset.
  • The green area covers one standard deviation from the mean.

4. Data Distribution by Feature

You can also zoom on distributions to understand what has changed.
To change the bins displayed, you can define custom options.

Report customization

You can set different Options for Data / Target drift to modify the existing components of the report. Use this to change the statistical tests used, define Dataset Drift conditions, or change histogram Bins.
You can also set Options for Quality Metrics to define the width of the confidence interval displayed for individual feature drift.
You can also select which components of the reports to display or choose to show the short version of the report: Select Widgets.
If you want to create a new plot or metric, you can Custom Widgets and Tabs.

When to use this report

Here are a few ideas on when to use the report:
  1. 1.
    In production: as early monitoring of model quality. In absence of ground truth labels, you can monitor for changes in the input data. Use it e.g. to decide when to retrain the model, apply business logic on top of the model output, or whether to act on predictions. You can combine it with monitoring model outputs using the Numerical or Categorical Target Drift report.
  2. 2.
    In production: to debug the model decay. Use the tool to explore how the input data has changed.
  3. 3.
    In A/B test or trial use. Detect training-serving skew and get the context to interpret test results.
  4. 4.
    Before deployment. Understand drift in the offline environment. Explore past shifts in the data to define retraining needs and monitoring strategies. Here is a blog about it.
  5. 5.
    To find useful features when building a model. You can also use the tool to compare feature distributions in different classes to surface the best discriminants.

JSON Profile

If you choose to generate a JSON profile, it will contain the following information:
{
"data_drift": {
"name": "data_drift",
"datetime": "datetime",
"data": {
"utility_columns": {
"date": null,
"id": null,
"target": null,
"prediction": null,
"drift_conf_level": value,
"drift_features_share": value,
"nbinsx": {
"feature_name": value,
"feature_name": value
},
"xbins": null
},
},
"cat_feature_names": [],
"num_feature_names": [],
"metrics": {
"feature_name" :{
"prod_small_hist": [
[],
[]
],
"ref_small_hist": [
[],
[]
],
"feature_type": "num",
"p_value": p_value
},
"n_features": value,
"n_drifted_features": value,
"share_drifted_features": value,
"dataset_drift": false
}
},
"timestamp": "timestamp"
}

Data Drift Dashboard Examples

  • Browse our example notebooks to see sample Reports.
You can also read the initial release blog.
Last modified 12d ago
Copy link
On this page
Summary
Requirements
How it works
How it looks
1. Data Drift Summary
2. Data Drift Table
3. Data Drift by Feature
4. Data Distribution by Feature
Report customization
When to use this report
JSON Profile
Data Drift Dashboard Examples