Prediction distribution monitoring in MLOps - Commands & Configuration

Introduction

Prediction distribution monitoring helps you watch how the results from your machine learning model change over time. It solves the problem of models becoming less accurate because the data they see in real life changes from what they were trained on.

When you want to check if your model's predictions are drifting away from what it learned during training

When you deploy a model and want to ensure it keeps making reliable predictions over weeks or months

When you want to detect sudden changes in the type of data your model is seeing, which might mean a problem

When you want to compare the distribution of new predictions to the original training predictions

When you want to alert your team if the model's prediction patterns change significantly

Commands

This command installs the Evidently library, which helps monitor prediction distributions easily.

Terminal

pip install evidently

Expected Output

Collecting evidently

  Downloading evidently-0.2.45-py3-none-any.whl (200 kB)

Installing collected packages: evidently

Successfully installed evidently-0.2.45

Runs a Python script that loads prediction data, compares current predictions to baseline, and prints a report on distribution changes.

Terminal

python monitor_predictions.py

Expected Output

Prediction distribution report:

- Kolmogorov-Smirnov test p-value: 0.03

- Warning: Prediction distribution has shifted significantly

- Mean prediction changed from 0.45 to 0.60

Key Concept

If you remember nothing else from this pattern, remember: monitoring prediction distributions helps catch when your model's outputs start to drift from what it learned, so you can fix problems early.

Code Example

MLOps

from evidently import ColumnMapping
from evidently.dashboard import Dashboard
from evidently.dashboard.tabs import DataDriftTab
import pandas as pd

# Load baseline and current prediction data
baseline = pd.read_csv('baseline_predictions.csv')
current = pd.read_csv('current_predictions.csv')

# Define which column holds predictions
column_mapping = ColumnMapping(prediction='prediction')

# Create a dashboard to check data drift in predictions
dashboard = Dashboard(tabs=[DataDriftTab()])
dashboard.calculate(baseline, current, column_mapping=column_mapping)

# Save the report as an HTML file
dashboard.save('prediction_drift_report.html')
print('Prediction drift report saved as prediction_drift_report.html')

Common Mistakes

Not comparing current predictions to a baseline distribution

Without a baseline, you cannot tell if the prediction distribution has changed or not.

Always save and use the original training or validation prediction distribution as a baseline for comparison.

Ignoring small but consistent changes in prediction distribution

Small changes can accumulate and degrade model performance over time if not addressed.

Set thresholds to detect even small shifts and monitor trends regularly.

Summary

Install the Evidently library to help monitor prediction distributions.

Run a script that compares current model predictions to a baseline to detect distribution changes.

Use a dashboard report to visualize and understand prediction drift over time.

Practice

1. What is the main purpose of prediction distribution monitoring in MLOps?

easy

A. To monitor the training data quality only

B. To track changes in the model's output predictions over time

C. To improve the speed of model training

D. To increase the size of the prediction dataset

Solution

1. Step 1: Understand prediction distribution monitoring

   It focuses on watching the outputs (predictions) of a model to detect changes or shifts.

2. Step 2: Differentiate from other monitoring types

   It is not about training data quality or training speed but about output behavior over time.

3. Final Answer:

   To track changes in the model's output predictions over time -> Option B

4. Quick Check:

   Prediction monitoring = track output changes [OK]

Hint: Focus on what is monitored: model outputs, not inputs or speed [OK]

Common Mistakes:

• Confusing prediction monitoring with data quality monitoring
• Thinking it speeds up training
• Assuming it increases dataset size

2. Which of the following is the correct way to calculate the distribution of predictions in Python using NumPy?

easy

A. np.sort(predictions, bins=10)

B. np.mean(predictions, bins=10)

C. np.sum(predictions, bins=10)

D. np.histogram(predictions, bins=10)

Solution

1. Step 1: Identify the function for distribution calculation

   NumPy's np.histogram calculates the frequency distribution of values in bins.

2. Step 2: Check other options

   np.mean calculates average, np.sum sums values, and np.sort sorts values, none calculate distribution.

3. Final Answer:

   np.histogram(predictions, bins=10) -> Option D

4. Quick Check:

   Distribution = histogram [OK]

Hint: Use np.histogram to get frequency counts in bins [OK]

Common Mistakes:

• Using mean or sum instead of histogram for distribution
• Trying to sort to get distribution
• Passing wrong arguments to functions

3. Given the following Python code snippet for monitoring prediction distribution, what will be the output?

import numpy as np
predictions = np.array([0.1, 0.4, 0.35, 0.8, 0.9])
hist, bins = np.histogram(predictions, bins=3)
print(hist)

medium

A. [3 1 1]

B. [1 2 2]

C. [2 1 2]

D. [2 2 1]

Solution

1. Step 1: Understand bin edges

   With bins=3, the range 0.1 to 0.9 is split into 3 equal parts: approx [0.1-0.4), [0.4-0.7), [0.7-1.0].

2. Step 2: Count predictions in each bin

   Bin 1: 0.1, 0.4 (0.4 is right edge, goes to next bin) -> 0.1 only -> 1 count Bin 2: 0.4, 0.35 -> 0.35 and 0.4 -> 2 counts Bin 3: 0.8, 0.9 -> 2 counts

3. Step 3: Correct bin counts

   Actually, np.histogram includes left edge, excludes right except last bin. So bins: [0.1,0.4), [0.4,0.7), [0.7,1.0] Values: 0.1 in bin1 0.35 in bin1 0.4 in bin2 0.8 in bin3 0.9 in bin3 Counts: bin1=2, bin2=1, bin3=2

4. Final Answer:

   [2 1 2] -> Option C

5. Quick Check:

   Histogram counts = [2,1,2] [OK]

Hint: Remember np.histogram includes left edge, excludes right edge except last bin [OK]

Common Mistakes:

• Miscounting values on bin edges
• Assuming bins include right edge
• Confusing bin counts order

4. You have this monitoring code snippet that throws an error:

import numpy as np
predictions = [0.2, 0.5, 0.7]
hist, bins = np.histogram(predictions, bins='five')
print(hist)

What is the cause of the error?

medium

A. The bins parameter must be an integer or sequence, not a string

B. The predictions list must be a NumPy array, not a list

C. The print statement syntax is incorrect

D. np.histogram does not accept more than 3 values

Solution

1. Step 1: Check bins parameter type

   np.histogram expects bins as an integer or a sequence of bin edges, not a string like 'five'.

2. Step 2: Verify other parts

   Predictions can be a list or array, print syntax is correct, and np.histogram accepts any length array.

3. Final Answer:

   The bins parameter must be an integer or sequence, not a string -> Option A

4. Quick Check:

   Bins must be int or list, not string [OK]

Hint: Bins must be number or list, never a string [OK]

Common Mistakes:

• Thinking list input causes error
• Blaming print syntax
• Assuming np.histogram limits input size

5. You want to detect if your model's prediction distribution has shifted significantly from the baseline. Which approach is best to implement in your monitoring pipeline?

hard

A. Calculate the KL divergence between baseline and current prediction distributions regularly

B. Only check if the average prediction value changes

C. Retrain the model every day regardless of prediction changes

D. Ignore distribution changes and focus on input data monitoring

Solution

1. Step 1: Understand distribution shift detection

   KL divergence measures how one distribution differs from another, ideal for detecting prediction shifts.

2. Step 2: Evaluate other options

   Checking only average misses distribution shape changes; retraining blindly wastes resources; ignoring prediction changes misses key signals.

3. Final Answer:

   Calculate the KL divergence between baseline and current prediction distributions regularly -> Option A

4. Quick Check:

   Use KL divergence for distribution shift detection [OK]

Hint: Use KL divergence to compare distributions, not just averages [OK]

Common Mistakes:

• Monitoring only average values
• Retraining without monitoring
• Ignoring prediction distribution shifts