Beyond Co-purchase Relation: Evolution of Complementary Recommendations at Allegro
Allegro.com deployed AlleCompanion, a production-scale retrieval framework for complementary product recommendations. The framework uses a category-constrained Two Tower architecture with a Category Adapter and a multi-source Complementary Categories Mapping called ComCat. ComCat integrates expert rules, human-in-the-loop feedback, and LLM-based reasoning to distill patterns from noisy co-purchase traffic.
Allegro introduced AlleCompanion to improve complementary product recommendations by distinguishing items that are merely bought together from those that truly work together. The system combines data-level filtering heuristics with a category-constrained Two Tower architecture. A Category Adapter guides the model in embedding space, while ComCat, a multi-source Complementary Categories Mapping, acts as a translational layer integrating expert rules, human-in-the-loop feedback, and LLM-based reasoning to create a maintainable and controllable solution.
The category-constrained Two Tower architecture with a Category Adapter constrains candidates within logically complementary boundaries in the embedding space. ComCat serves as a translational layer that distills meaningful patterns from noisy traffic, combining expert rules, human feedback, and LLM-based reasoning. This approach addresses the intrinsic noise in large-scale co-purchase traffic and improves semantic compatibility modeling.
E-commerce platforms are moving beyond simple co-purchase signals to semantic compatibility for basket building. The integration of LLM-based reasoning with human-in-the-loop feedback and expert rules indicates a trend toward hybrid, maintainable recommendation systems that balance automation with control.
Improved complementary recommendations can increase basket size and customer satisfaction by suggesting items that genuinely work together, potentially boosting revenue per session and reducing irrelevant recommendations. The maintainable and controllable nature of ComCat may lower operational costs and improve trust in the recommendation system.
Observable next signals include publication of detailed performance metrics, adoption of similar category-constrained architectures by other e-commerce platforms, and further integration of LLM-based reasoning in production recommendation systems. Potential expansion of ComCat to other product domains or marketplaces may follow.