Abstract
Identifying item relationships is fundamental to e-commerce recommendation systems. This paper introduces EMRIGCN, a novel multi-task Graph Convolutional Network designed to simultaneously infer substitutable and complementary items. By employing a two-level integration strategy using Multi-gate Mixture-of-Experts (MMoE) and incorporating relationship strength via an auxiliary loss, EMRIGCN addresses the limitations of existing models in handling mutual influence and data sparsity. Experimental results on public and industrial datasets demonstrate significant improvements over state-of-the-art baselines, particularly in cold-start scenarios and real-world production environments.
Introduction
In the competitive landscape of e-commerce, the "secret sauce" of effective recommendation systems lies in understanding how items relate to one another. These relationships primarily fall into two categories: substitutes (items that are interchangeable, such as different brands of the same product) and complements (items typically purchased together, such as bread and milk). Accurately identifying these links is crucial for optimizing user experience and driving business metrics like Revenue Per Thousand Impressions (RPM).
The Challenge of Item Relationships
Existing Graph Neural Networks (GNNs) often treat substitute and complement relationships in isolation, failing to capture the mutual influence between them. For instance, knowing an item's substitutes can provide valuable context for predicting its potential complements. Furthermore, most models treat these relationships as binary (0 or 1), ignoring the varying "strength" of links—bread and milk share a much stronger complementary bond than bread and luncheon meat.
Traditional models like DecGCN struggle particularly in "cold-start" scenarios where an item might have known substitutes but no recorded complements in the training data. Without a mechanism to transfer knowledge between relationship types, these models provide suboptimal recommendations for new or niche products.
Methodology: EMRIGCN
To address these challenges, we propose the Enhanced Multi-Relationships Integration Graph Convolutional Network (EMRIGCN). This model treats relationship inference as a link prediction task on a heterogeneous graph, utilizing two key innovations:
1. Two-Level Integration Module
EMRIGCN leverages a Multi-gate Mixture-of-Experts (MMoE) architecture at two distinct stages to capture the interplay between relationship types:
- Feature Integration (Low-level): A shared encoder uses MMoE to extract both relationship-specific and shared features from raw item attributes (e.g., category, brand) at the initial encoding stage.
- Structure Integration (High-level): During the GCN's message-passing phase, the model aggregates information from both substitute and complement neighbors. An MMoE layer then adaptively fuses this information, allowing the model to "transfer knowledge" across relationship types.
2. Relationship Enhancement via Auxiliary Loss
To move beyond binary labels and incorporate relationship strength, we introduce an auxiliary triplet loss. This function forces the model to rank "strong" related items closer in the embedding space than "weak" related items, using co-occurrence frequencies (e.g., how often items appear in the same order) as a proxy for strength.
Experimental Evaluation
The model was rigorously evaluated on four Amazon public datasets (Beauty, Electronics, Cellphones, and Clothing) and a large-scale industrial dataset from Meituan Maicai, a Chinese online supermarket.
Key Results
- State-of-the-Art Performance: EMRIGCN consistently outperformed established baselines such as PinSage, R-GCN, and DecGCN. On the Amazon Beauty dataset, it improved Recall@30 for complements by approximately 1.5% over the strongest baseline.
- Cold-Start Robustness: In scenarios with fewer than four neighbors, EMRIGCN demonstrated a significant lead, outperforming PinSage by over 15% in Recall@1 on the Beauty dataset. This highlights the effectiveness of the MMoE-based knowledge transfer.
- Real-World Impact: A/B testing on the Meituan Maicai platform revealed a 15.3% increase in Visit Buy Rate (VBR) and a 15.34% increase in RPM, validating the model's utility in a live production environment.
Discussion and Future Outlook
The success of EMRIGCN underscores the importance of integrating multiple relationship types and accounting for relationship strength. By using shared experts, the model automatically determines which features are universal and which are relationship-specific, providing a robust solution for data-sparse scenarios.
However, the computational complexity of applying MMoE layers at every convolution step warrants further investigation regarding real-time inference latency at scale. Future research could also extend this framework to include other relationship types, such as upselling or brand-loyalty links, and explore more nuanced definitions of relationship strength based on user reviews or return rates.
Conclusion
EMRIGCN represents a significant step forward in item relationship inference. By holistically modeling the synergy between substitutes and complements, it provides a more accurate and nuanced understanding of e-commerce catalogs, leading to measurable improvements in both recommendation quality and business performance.
Summary updated