Personalizer Overview

Reinforcement learning service for personalized content recommendations based on user behavior and context

Personalizer

Personalizer has been retired. The service is no longer available for new projects. Existing applications should migrate to alternative solutions.

Azure AI Personalizer was a reinforcement learning service that helped applications choose the best content or actions to show users based on their real-time behavior and context. The service learned from user feedback to optimize recommendations over time.

What Was Personalizer?

Personalizer used reinforcement learning to:

• Rank actions: Determine the best action for a given context
• Learn from feedback: Improve decisions based on rewards
• Personalize at scale: Optimize for all users collectively
• Adapt in real-time: Update models continuously

This documentation is provided for historical reference and migration planning only.

How It Worked

Personalizer used two primary APIs:

Rank API

Determine the best action for current context:

from azure.cognitiveservices.personalizer import PersonalizerClient
from azure.cognitiveservices.personalizer.models import RankRequest

client = PersonalizerClient(endpoint, credentials)

# Define actions (content options)
actions = [
    {"id": "article1", "features": [{"topic": "tech", "length": "short"}]},
    {"id": "article2", "features": [{"topic": "sports", "length": "long"}]},
    {"id": "article3", "features": [{"topic": "news", "length": "medium"}]}
]

# Define context (user and situation)
context = [
    {"user": {"location": "US", "device": "mobile"}},
    {"time": {"dayOfWeek": "Monday", "hour": 14}}
]

# Get ranking
rank_request = RankRequest(
    actions=actions,
    context_features=context,
    event_id="unique-event-id"
)

response = client.rank(rank_request)

# Use top-ranked action
top_action = response.ranking[0]
print(f"Show content: {top_action.id}")
print(f"Probability: {top_action.probability}")

Reward API

Provide feedback on the recommendation:

# User clicked on the article (positive outcome)
client.events.reward(
    event_id="unique-event-id",
    value=1.0  # Reward score: 0 (bad) to 1 (good)
)

# User ignored the article (negative outcome)
client.events.reward(
    event_id="unique-event-id",
    value=0.0
)

# Partial reward for incomplete action
client.events.reward(
    event_id="unique-event-id",
    value=0.5
)

Key Concepts

Actions

Content items or choices to rank:

• ID: Unique identifier for the action
• Features: Characteristics describing the action
• Example: Articles, products, ads, recommendations

{
  "id": "product-123",
  "features": [
    {"category": "electronics", "price": "high", "brand": "premium"}
  ]
}

Context

Information about the current situation:

• User features: Demographics, preferences, history
• Environment features: Time, device, location
• Session features: Page type, previous actions

[
  {"user": {"age": "25-34", "interests": ["tech", "gaming"]}},
  {"device": {"type": "mobile", "os": "iOS"}},
  {"time": {"hour": 20, "dayOfWeek": "Friday"}}
]

Rewards

Feedback indicating outcome quality:

• Range: 0 (worst) to 1 (best)
• Timing: Real-time or delayed
• Meaning: Defined by your business objectives

Example Reward Logic:

User Behavior	Reward Score	Reasoning
Clicked and purchased	1.0	Best outcome
Clicked, viewed 90%+	0.8	Strong engagement
Clicked, viewed 30%+	0.5	Moderate interest
Clicked, bounced quickly	0.2	Poor match
Did not click	0.0	Not relevant

Learning Modes

Apprentice Mode

Safe training mode for new models:

• Learn from your existing logic
• No impact on user experience
• Validate features and configuration
• Returns your baseline action
• Build confidence before going live

# In Apprentice mode, baseline action is returned
rank_request = RankRequest(
    actions=actions,
    context_features=context,
    baseline_action="default-content"  # Your current logic
)

Online Mode

Production mode with active learning:

• Returns best predicted action
• Explores alternative actions
• Learns from all feedback
• Continuously improves
• Optimizes for reward maximization

Use Cases

Personalizer was used for:

Content Personalization

• News article recommendations
• Video suggestions
• Product recommendations
• Ad placement optimization
• Email content selection

User Experience

• Homepage layout personalization
• Feature highlighting
• Navigation customization
• Search result ranking
• Notification timing

E-commerce

• Product recommendations
• Promotional offers
• Pricing strategies
• Bundle suggestions
• Upsell opportunities

Communication

• Push notification timing
• Email subject lines
• Message content selection
• Communication channel choice

Features and Configuration

Feature Engineering

Design effective features:

• Categorical: Device type, location, category
• Numerical: Price, rating, popularity
• Boolean: Is premium, has discount, in stock
• Text: Tags, keywords (as categorical)

Best Practices:

• Use 5-50 features per action/context
• Include diverse feature types
• Avoid highly correlated features
• Update features as they change
• Test feature importance

Model Configuration

• Update frequency: How often model trains
• Exploration percentage: Random action probability
• Reward wait time: Delay for reward signal
• Default reward: Reward when none provided

Evaluation

• Offline evaluation: Compare strategies on historical data
• Online evaluation: A/B test in production
• Counterfactual evaluation: Estimate alternative policies

Migration Alternatives

With Personalizer retired, consider:

Azure AI Solutions

• Azure OpenAI: For LLM-based recommendations
• Azure Machine Learning: Custom recommendation models
• Azure Applied AI: Pre-built AI solutions

Recommendation Systems

• Collaborative filtering: User-item matrix factorization
• Content-based filtering: Feature-based matching
• Hybrid approaches: Combine multiple techniques

Third-Party Services

• Amazon Personalize
• Google Recommendations AI
• Algolia Recommend
• Custom ML solutions

Best Practices (Historical)

For reference, best practices included:

Data Requirements

• Minimum 1,000 events per day for learning
• Consistent feature schema
• Quality reward signals
• Sufficient action variety (50 max recommended)

Feature Design

• Use relevant, non-redundant features
• Include both action and context features
• Update features when they change
• Test feature importance with evaluations

Reward Function

• Align with business objectives
• Provide timely feedback
• Use consistent scale (0-1)
• Consider delayed rewards for complex goals

Testing

• Start in Apprentice mode
• Run offline evaluations
• A/B test before full deployment
• Monitor metrics continuously

SDK Support (Historical)

Personalizer supported:

Python

pip install azure-cognitiveservices-personalizer

C#

dotnet add package Microsoft.Azure.CognitiveServices.Personalizer

JavaScript

npm install @azure/cognitiveservices-personalizer

Java

Maven package available

Migration Guidance

1. Assess Current Usage

   Document your Personalizer implementation, features, and metrics

2. Evaluate Alternatives

   Research replacement solutions based on your requirements

3. Export Data

   Save historical event and model data before decommissioning

4. Implement New Solution

   Build and test replacement recommendation system

5. Migrate Traffic

   Gradually transition users to new system

6. Decommission

   Remove Personalizer integration after validation

Resources

• Retirement Announcement
• Historical Documentation
• Code Samples
• Migration Guide

Support

For existing Personalizer implementations:

• Review retirement timeline and deadlines
• Plan migration strategy
• Export critical data
• Implement alternative solutions
• Contact Azure support for migration assistance