Increasing purchases by 9%

SidelineSwap is a leading online marketplace for new and used sporting equipment, where athletes can buy, sell, and trade their gear. The platform serves a diverse community of sports enthusiasts across multiple categories including hockey, baseball, golf, and other sporting equipment. With a large inventory of items and a growing user base, SidelineSwap has become the go-to destination for athletes looking to find the right gear at the right price.

In just 6 weeks, Shaped transformed SidelineSwap's recommendation system from a static, one-size-fits-all approach to a sophisticated real-time personalization engine, delivering:

• 9% uplift in both clicks and purchases
• 100% accuracy in size-specific recommendations
• 65% → 100% product attribute coverage through intelligent feature engineering
• Real-time personalization across multiple recommendation surfaces

The success stemmed from Shaped's unique approach to solving the "size problem" in sports equipment recommendations - ensuring users viewing size 13 hockey skates only see other size 13 equipment. This was achieved through a hybrid architecture combining ML-based personalization with deterministic filtering, backed by sophisticated feature engineering to handle sparse product data.

This case study details how we built a system that not only understands product relationships but also maintains perfect size matching - a critical requirement for sports equipment recommendations that typical recommendation systems struggle to handle.

SidelineSwap faced significant challenges with their recommendation system. Their existing solution was basic and static, lacking true personalization capabilities. Two critical problems emerged:

1. Content Understanding: The system struggled to make meaningful connections between related sporting equipment categories and user preferences.
2. Size Personalization: A crucial challenge was the inability to factor in sizing when making recommendations. For example, users viewing size 13 hockey skates would receive recommendations for skates in various sizes, leading to poor user experience and reduced conversion rates.

The complexity was amplified by data quality issues, with only 65% of products having detailed attribute information, creating significant sparsity in the dataset.

Phase 1: Building the Foundation

The implementation began with developing a base recommender model using static datasets, focusing initially on product categorization (e.g., hockey skates, hockey sticks). This established the fundamental personalization layer based on sport-specific categories and user interactions.

Phase 2: Real-Time Integration

The system was then evolved to handle real-time data by integrating:

• Production PostgreSQL database
• S3 items dataset
• Amplitude event stream

Phase 3: Advanced Feature Engineering

A sophisticated feature engineering process was implemented to address the sizing and categorization challenges:

1. Key Detail Identification:
   • Developed a system to identify the most critical sizing-related attribute for each product category
   • Examples:
     • Baseball bats → length
     • Golf drivers → handedness
     • Hockey skates → shoe size
2. Handling Data Sparsity:
   • Created a complex SQL-based hierarchy ordering system
   • Implemented a fallback mechanism to find similar products with complete details
   • Used feature engineering to fill data gaps through attribute inference

Phase 4: Hybrid Modeling Approach

The solution required a unique combination of:

• Machine learning for category-based recommendations
• Deterministic filtering for size matching
• Dual modeling of key details as both categories and text
• Content similarity understanding through text encoding
• Building embedding spaces to identify hyper-similar items

Technical Implementation Details

• Implementation timeline: Approximately 6 weeks to initial deployment
• Event weighting system:
  • Clicks: Base weight (1)
  • Favorites: Medium weight (5)
  • Purchases: Highest weight (10)
• Used Shaped's built-in content similarity capabilities for text encoding and embedding space construction
• Implemented through SQL interfaces for rapid iteration on feature engineering
  

Immediate Outcomes

• 9% uplift in both clicks and purchases compared to the previous system
• Successfully achieved 100% accuracy in size-specific recommendations
• Improved data quality awareness, helping SidelineSwap identify and address data maintenance issues

Extended Implementation

The success led to additional use cases:

1. Complementary item recommendations (e.g., baseball glove → baseball bat)
2. Personalized category rankings for homepage customization
3. Development of a new model version with:
   • Improved text and image content understanding
   • Enhanced processing speed
   • New user onboarding integration, balancing initial interests with growing personalization

1. Feature Modeling Trade-offs:
   • Category-based modeling impacts different dynamics compared to text-based modeling
   • Hybrid approaches often yield optimal results
   • Shaped's flexibility allows easy experimentation with different modeling approaches
2. Infrastructure Considerations:
   • Real-time data processing requirements
   • Balance between ML-driven personalization and deterministic filtering
   • Importance of rapid iteration capabilities in feature engineering
3. Data Quality Impact:
   • Critical importance of comprehensive product attributes
   • Need for robust fallback mechanisms in sparse data scenarios
   • Value of SQL-based feature engineering for rapid prototyping

The project demonstrates how Shaped's platform enables rapid development of sophisticated recommendation systems, handling both the technical complexity of real-time personalization and the practical challenges of imperfect data in production environments.