Why Value Modeling?

Traditional recommendation systems often fall short when it comes to optimizing for complex, nuanced business goals. Many machine learning-based ranking models prioritize broad engagement metrics like clicks and conversions but lack the ability to integrate business-specific objectives such as lifetime value, inventory optimization, or compliance constraints. Shaped Value Modeling changes that by enabling you to:

1. A Control Panel for Your Business Objectives

Traditional models optimize for a single objective and require lengthy retraining to incorporate new ranking logic. Value Modeling acts as a control panel, allowing teams to fine-tune rankings dynamically, optimizing engagement, business metrics, and quality scores without rebuilding models.

2. Open the Black Box and Take Control

ML models often lack visibility and flexibility, making it difficult to understand why certain recommendations rank higher than others. Value Modeling provides direct observability, letting teams define precise ranking logic using interpretable expressions.

3. Experiment and Iterate Without Retraining

No more waiting for lengthy training cycles. Adjust ranking weights, test different scoring strategies, and refine recommendations in real time. Business objectives change frequently—your ranking model should adapt just as fast.

4. Blend Multiple Models for Smarter Ranking

Instead of using one model to optimize for all objectives, Value Modeling allows you to train separate models for clicks, purchases, or quality—and then combine them dynamically.

value_model: 0.5 * click_model + 0.5 * purchase_model

How Shaped Value Modeling Works

Shaped Value Modeling utilizes the familiar Jinja templating framework, enabling you to define your scoring logic using intuitive Python expressions. This provides unmatched flexibility and allows for seamless integration with your existing data and workflows.

Supported Expressions:

• Python code: Leverage the full power of Python to express complex scoring logic.
• Jinja filters: Simplify common value model use cases with built-in Jinja filters.
• Math functions: Access the full range of Python's math library directly within your value model.
• Built-in functions: Utilize helpful functions like len, sum, max, min, abs, and now_seconds().
• Feature access: Incorporate user (user), item (item), and interaction (user.recent_interactions) features directly into your scoring logic.

Getting Started with Shaped Value Modeling

Defining a value model with Shaped is simple. Within your Score Ensemble Policy configuration, use the value_model parameter to specify your scoring expression:

{
  "user_id": "83NSLX",
  "config": {
    "value_model": 0.3 * lightgbm + 0.7 * bert4rec
  }
}

Real-World Use Cases

1. E-Commerce Personalization

Optimize recommendations by balancing engagement, purchase intent, and product diversity.

0.3 * click_model + 
0.6 * purchase_model + 
0.1 * item.diversity_score

2. News & Content Platforms

Dynamically decay rankings for older content while boosting high-quality journalism.

exp(-0.28 * (now_seconds() - item.created_at) / 3600) * 
(0.4 * item.is_author_verified + 
 0.6 * engagement_model - 
 0.4 * clickbait_model)

3. Marketplace Local Ranking

Dynamically mix personalized scores with a geo-location penalizer to connect users with nearby items.

exp(-0.015 * (2 * 6371 * 
asin(sqrt(0.5 - cos(radians(item.lat - user.lat)) / 2 + 
cos(radians(user.lat)) * cos(radians(item.lat)) * 
(1 - cos(radians(item.long - user.long)))/2)))) * 
(0.8 * purchase_model + 
 0.2 * favorite_model)

4. Social Media Ranking Systems

Switch between personalized & trending models dynamically.

((user.recent_interactions.label|select('gt', 0)|list|count) >= 3) * 
(0.22 * like_model + 
 0.47 * comment_model + 
 0.31 * share_model) + 
((user.recent_interactions.label|select('gt', 0)|list|count) < 3) * 
recently_popular

Value Modeling empowers you to take full control of your ranking logic, ensuring that your recommendations align with your business goals in real time. Whether you're optimizing for engagement, revenue, content quality, or a combination of objectives, this flexible framework eliminates the trade-offs of traditional black-box models.

Ready to see it in action? Book a demo today and discover how Shaped can help you build smarter, more adaptable recommendation systems that drive real business impact.

‍

‍