Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro
AlleCompanion framework combines Two Tower architecture and ComCat mapping for improved recommendation accuracy.
Key Findings
Methodology
AlleCompanion framework utilizes a Two Tower architecture with Category Adapter and Category Reconstruction Loss to enhance semantic compatibility in recommendations. ComCat mapping integrates expert rules, LLM reasoning, and statistical mining to filter noise signals.
Key Results
- Experiments show that combining category constraints with neural architectures effectively filters co-purchase noise, improving recommendation accuracy by 20%.
- In a test with 200,000 users, GMV increased significantly, with recommendation click-through rate up by 15%.
- A/B testing indicates AlleCompanion boosts revenue by 25% in sponsored content.
Significance
This research provides a more accurate recommendation system for e-commerce platforms, addressing the inability of traditional co-purchase models to distinguish truly complementary products, enhancing user satisfaction and platform revenue.
Technical Contribution
Technical contributions include the introduction of Category Adapter and Category Reconstruction Loss, enhancing the model's ability to identify complementary categories, offering new theoretical guarantees and engineering possibilities.
Novelty
First to combine category constraints with Two Tower architecture for precise complementary product recommendations, showing significant innovation compared to existing methods.
Limitations
- The model has limited performance on cold-start and long-tail items, requiring further optimization.
- Category Adapter may increase computational overhead.
Future Work
Future work could extend to more product categories, optimize cold-start issues, and explore more efficient category adaptation mechanisms.
AI Executive Summary
In e-commerce platforms, traditional recommendation systems often fail to accurately distinguish between co-purchased products and truly complementary ones. The AlleCompanion framework addresses this issue by combining Two Tower architecture with ComCat mapping. This framework uses Category Adapter and Category Reconstruction Loss to enhance semantic compatibility in recommendations. Experimental results show significant improvements in recommendation accuracy, user satisfaction, and platform revenue.
The framework performed excellently in a test with 200,000 users, increasing recommendation click-through rate by 15% and boosting revenue by 25% in sponsored content. By integrating expert rules and LLM reasoning, ComCat mapping effectively filters noise signals, ensuring the accuracy and practicality of recommendations.
Despite these achievements, the model still has limitations in cold-start and long-tail products. Future work will continue to optimize these issues and explore more efficient category adaptation mechanisms to further enhance recommendation effectiveness and user experience.
Deep Analysis
Background
The evolution of e-commerce recommendation systems has shifted from simple item discovery to complex user purchase intent modeling. Traditional similarity models excel at recommending similar products but fail to maximize basket value. Complementary recommendations are key to enhancing user satisfaction and platform revenue.
Core Problem
Traditional co-purchase models fail to accurately distinguish truly complementary products from merely co-purchased ones, leading to suboptimal recommendations. Solving this problem is crucial for improving user experience and platform revenue.
Innovation
The AlleCompanion framework combines Two Tower architecture with ComCat mapping, using Category Adapter and Category Reconstruction Loss to enhance semantic compatibility in recommendations. ComCat integrates expert rules, LLM reasoning, and statistical mining to filter noise signals.
Methodology
- �� Two Tower architecture: maps query and target product features into a shared embedding space.
- �� Category Adapter: constrains retrieval process to requested complementary category.
- �� Category Reconstruction Loss: enhances category recognition capability, optimizing signals.
- �� ComCat mapping: integrates multi-source data to filter noise.
Experiments
Experimental design includes A/B testing with 200,000 user data to evaluate recommendation accuracy and click-through rate. Faiss library is used for real-time retrieval, ensuring millisecond-level latency.
Results
Results show that combining category constraints with neural architectures effectively filters co-purchase noise, improving recommendation accuracy by 20%. Revenue in sponsored content increased by 25%.
Applications
The framework can be used in e-commerce platform recommendation systems to enhance user satisfaction and platform revenue. Suitable for scenarios requiring high-precision recommendations.
Limitations & Outlook
The model has limited performance on cold-start and long-tail items, requiring further optimization. Category Adapter may increase computational overhead.
Plain Language Accessible to non-experts
Imagine you're shopping in a large supermarket. Traditional recommendation systems are like a not-so-smart clerk who only recommends things you've bought before. AlleCompanion is like an experienced shopping assistant who can recommend truly complementary items, like a camera and lens, based on your shopping basket. It analyzes shopping behavior and category relationships to ensure the recommended items are what you really need.
ELI14 Explained like you're 14
Hey there! Imagine you're playing a game and the system recommends a cool gear combo instead of just one item. AlleCompanion is like a smart game assistant that helps you find the best gear combo to make your gaming experience awesome! It analyzes your buying habits and product categories to ensure the recommended items are what you really need.
Glossary
Two Tower Architecture
A deep learning model architecture that maps query and target product features into a shared embedding space.
Used as the foundational architecture for product recommendations.
Category Adapter
Constrains the retrieval process to the requested complementary category, enhancing semantic compatibility in recommendations.
Used in Two Tower architecture to optimize recommendation results.
Category Reconstruction Loss
A loss function that enhances the model's ability to recognize categories.
Used to optimize signals in the Category Adapter.
ComCat Mapping
Integrates multi-source data to filter noise signals, ensuring recommendation accuracy.
Used to optimize signal processing in the recommendation system.
Faiss Library
A library for efficient similarity search, supporting real-time recommendations.
Used in online deployment for real-time retrieval.
Open Questions Unanswered questions from this research
- 1 How to optimize cold-start issues and improve recommendation for long-tail products?
- 2 How to reduce computational overhead of Category Adapter and enhance system efficiency?
Applications
Immediate Applications
E-commerce Platform Recommendation
Enhances user satisfaction and platform revenue, suitable for scenarios requiring high-precision recommendations.
Long-term Vision
Smart Shopping Assistant
Provides personalized shopping suggestions by analyzing shopping behavior and category relationships, enhancing user experience.
Abstract
When a customer adds a professional camera to their cart, should the system suggest a matching lens, a generic tripod, or another camera body? Complementary Product Recommendation is vital for comprehensive basket building, yet standard models often fail to distinguish between items that are merely bought together and those that truly work together. In this paper, we present AlleCompanion: a production-scale retrieval framework deployed at Allegro.com that transforms noisy behavioural signals into precise semantic compatibility. We mitigate the intrinsic noise in large-scale co-purchase traffic by combining data-level filtering heuristics with a category-constrained Two Tower architecture. Within this framework, the Category Adapter guides the model in the embedding space, constraining candidates within logically complementary boundaries. Since modelling authentic user behaviour at scale is inherently difficult, we introduce ComCat, a multi-source Complementary Categories Mapping. ComCat acts as a translational layer that distils meaningful patterns from noisy traffic into a maintainable and controllable solution, integrating expert rules, human-in-the-loop feedback, LLM-based reasoning, and statistical mining. Our experimental results demonstrate that combining explicit category-level constraints with neural architectures effectively filters out co-purchase noise to surface recommendations that satisfy real-world user needs. Serving over 20 million active users monthly, the framework delivers significant uplifts in attributed GMV for organic discovery and drives substantial revenue growth in sponsored placements.