GNNs Power Zalando's Landing Page Personalization

Unlocking Engagement with GNNs

Mariia Bulycheva's presentation at InfoQ Dev Summit Munich provides a compelling case study on leveraging Graph Neural Networks (GNNs) to enhance Zalando's landing page recommender system. The key insight is the transition from traditional deep learning, which struggled with optimizing for long-term metrics like user retention and purchases, to GNNs, which are inherently better suited for modeling complex user-content interactions and higher-order relationships. The innovative aspect lies in the detailed explanation of converting tabular user logs into a heterogeneous graph, the intricacies of the 'message passing' training paradigm, and crucially, the practical challenges encountered, such as graph data leakage and inference latency. The proposed solution of a hybrid architecture to address latency and deliver contextual embeddings to downstream models is a significant takeaway for practitioners aiming for production-ready GNN deployments.

One notable limitation or concern, though implicitly addressed by the hybrid architecture, is the inherent complexity and computational cost associated with training and deploying GNNs at scale. The presentation touches upon this by discussing the need for careful data preparation, including creating separate train and test graphs to avoid leakage, and the utilization of libraries like Deep Graph Library and PyTorch Geometric. However, the exact trade-offs in terms of infrastructure, engineering effort, and potential for model drift in a dynamic e-commerce environment would be valuable additions. The benefits of this approach are clear for e-commerce platforms seeking deeper user understanding and more personalized recommendations, but also for any domain where complex relational data can be modeled as a graph, such as social networks, knowledge graphs, or fraud detection systems. The technical implications are substantial, pushing beyond simple feature embeddings to capture contextual relationships, thereby enabling more sophisticated and effective recommendation engines.

Key Points

• Transitioned from classic deep learning to Graph Neural Networks (GNNs) to improve Zalando's landing page recommender system for longer-term engagement metrics.
• Reimagined recommendation as a graph problem, with users and content as nodes and interactions (views, clicks) as edges.
• Detailed the process of converting tabular user logs into a heterogeneous graph, including node features derived from image embeddings of purchased/associated items.
• Explained the 'message passing' training process and highlighted critical pitfalls like graph data leakage, emphasizing the need for fully disconnected train and test graphs.
• Addressed inference latency challenges with a hybrid architecture that delivers contextual embeddings to a downstream model.

---

📖 Source: Presentation: Reimagining Platform Engagement with Graph Neural Networks