Drift Detection

When ML models are deployed in production, sometimes even minor changes in a data distribution can adversely affect the performance of ML models. When the input data distribution shifts then prediction quality can drop. It is important to track this drift. This demo is based on the mixed-type tabular data drift detection method in the alibi detect project for tabular datasets.

Here we will :

• Launch an income classifier model based on demographic features from a 1996 US census. The data instances contain a person’s characteristics like age, marital status or education while the label represents whether the person makes more or less than $50k per year.

• Setup a mixed-type tabular data drift detector for this particular model.

• Make a batch of predictions over time.

• Track the drift metrics in the Monitoring dashboard.

This demo requires Knative installation on the cluster as the drift detector will be installed as a kservice. See Knative installation instructions for necessary setup required.

Register an income classifier model

Register a pre-trained income classifier SKLearn model.

1. In the Model Catalog page, click Register a new model:

   

   "Register a new model" button on the Model Catalog page
2. In the Register New Model wizard, enter the following information, then click Register Model:

   • Model Name: income-classifier

   • URI: gs://seldon-models/scv2/samples/mlserver_1.6.0/income-sklearn/classifier/

   • Artifact Type: SciKit Learn

   • Version: v1

   

   Model configuration wizard

Configure predictions schema for classifier

Edit the model metadata to update the prediction schema for the model. The prediction schema is a generic schema structure for machine learning model predictions. It is a definition of feature inputs and output targets from the model prediction. Learn more about the predictions schema at the ML Predictions Schema open source repository. Use the income classifier model predictions schema income-classifier-prediction-schema.json to edit and save the model level metadata.

4KB

income-classifier-prediction-schema.json

1. Click the model income-classifier that you registered.

   

   Select "income-classifier" model on the Model Catalog page
2. Click Edit Metadata to update the Prediction schema field associated with the model using the contents of prediction schema income-classifier-prediction-schema.json.

   

   Model's metadata wizard
3. Click Save Metadata.

Launch a Seldon Deployment

Deploy the income classifier model from the catalog into an appropriate namespace

1. In the Model catalog, select Deploy of the Action dropdown.

   

   "Deploy" action in the dropdown on the Model Catalog page
2. Enter the deployment details in the deployment creation wizard and click Next:

   • Name: income-classifier

   • Namespace: seldon

   • Type: Seldon Deployment

   

   Income classifier deployment details
3. The predictor details should already be filled in from the model catalog except the Model Name field. Use income for the Model Name field.

   • Model Name: income

   The Model Name is linked to the name described in the model-settings.json file, located in the Google Cloud Storage location. Changing the name in the JSON file would also require changing the Model Name, and vice versa.

Income classifier deployment predictor

1. Click Next for the remaining steps, then click Launch.

Add a Drift Detector

From the deployment overview page, select your deployment to enter the deployment dashboard. Inside the deployment dashboard, add a drift detector with by clicking the Add button within the Drift Detection widget.

Expand to see drift detector creation

Enter the following parameters in the modal popup which appears, to configure the detector:

• Model Name: income-drift.

• Storage URI: (For public Google Buckets, secret field is optional)

  gs://seldon-models/scv2/samples/mlserver_1.6.0/income-sklearn/drift-detector

• Reply URL: Leave as the default value, shown below, unless you are using a custom installation, please change this parameter according to your installation

  http://seldon-request-logger.seldon-logs

• Minimum Batch Size: 200

• Drift Type: Feature

Then, click CREATE DETECTOR to complete the setup.

Configure predictions schema for detector

As per the income classifier model, use the same model predictions schema income-classifier-prediction-schema.json to edit and save the model level metadata for drift detector.

1. Click on the vertical ellipses “⋮” icon for the drift detector you have just registered.

   

   select detector
2. Click the Configure Metadata option to update the prediction schema associated with the model

3. Paste the downloaded income-classifier-prediction-schema.json, name the model income-drift and click Save Metadata.

   

   configure prediction schema

Run Batch Predictions

1. From the deployment dashboard, click on Batch Jobs. Run a batch prediction job using the Open Inference Protocol (OIP) payload format text predictions data file data.txt.

471KB

data.txt

This file has 4000 individual data points and based on our drift detector configuration, drift will be detected for a batch every 200 points. The distribution of the data in the first half section is the same as the distribution of the reference data the drift detector was configured with and the second half section of the data should be different to observe drift.

1. Upload the data to a bucket store of your choice. This demo will use MinIO and store the data at bucket path minio://income-batch-data/data.txt . Do not forget to configure your storage access credentials secret - we have it as minio-bucket-envvars here. Refer to the batch request demo for an example of how this can be done via the minio browser.

2. Running a batch job with the configuration below. This runs an offline job that makes a prediction request for a batch of 200 rows in the file at minio://income-batch-data/data.txt every 5 seconds:

Input Data Location: minio://income-batch-data/data.txt
Output Data Location: minio://income-batch-data/output-{{workflow.name}}.txt
Number of Workers: 1
Number of Retries: 3
Batch Size: 200
Minimum Batch Wait Interval (sec): 5
Method: Predict
Transport Protocol: REST
Input Data Type: Open Inference Protocol (OIP)
Storage Secret Name: minio-bucket-envvars

Monitor Drift Detection Metrics

Under the Monitor section of your deployment navigation, on the Drift Detection Tab, you can see a timeline of drift detection metrics.

The drift dashboard showcases 2 types of metrics graphs:

1. P-value score over time

   • 1a. Zoomed in view, focusing on features that have drifted, i.e. features that have a p-value score of less than the threshold.

   

   p-values drift metrics zoomed in

   • 2b. Zoomed out view, showing all features.

   

   p-values drift metrics zoomed out
2. Distance score over time.

distance score metrics

Note that for both drift metrics graphs, the starting batches do not drift and are marked by an empty O symbol, while the later batches do drift and are marked by a filled O symbol.

Monitor Drift Detection Alerts

If you have alerting configured you should see a notification about the drift

alert notification

with further details present on the alerting log

alert notification

Data drift and reference distributions comparison

To further analyse prediction data drift, you can also switch to the feature distribution tab to compare predictions to reference data distribution. See feature distribution monitoring demo for setup details.

Upload the income classifier reference dataset drift-reference-v2.csv as the reference data to monitor data drift in terms of feature distributions.

28KB

drift-reference-v2.csv

Once reference data is available, you can compare the distributions of the prediction data to the reference data.

You can see when reference data is available by checking the button on the top left of the Distributions dashboard. If it is not clickable and displays Reference data available, then reference data is available.

Reference data available

For each feature, you can click on Toggle reference data to view reference data side by side.

Monitor distributions

We will see that the drifted data has lower education individuals that were not in the reference data.

Troubleshooting

If you experience issues with this demo, see the troubleshooting docs and also the Knative or Elasticsearch sections.