ContextGNN: Beyond Two-Tower Recommendation Systems

TL;DR

ContextGNN presents a hybrid graph neural network architecture for temporal, heterogeneous recommendations that fuses pair-wise representations in a user’s local subgraph with a lightweight two-tower model for distant items. The model learns a user-specific fusion to adaptively weight local versus distant signals, enabling effective recommendations for both familiar and exploratory items. Empirical results on RelBench demonstrate significant gains over both purely pair-wise and purely two-tower baselines, with notable robustness to locality variations, and the approach provides scalable training via a sampled-softmax objective. The work advances relational deep learning by offering a unified, end-to-end framework that integrates local, context-rich patterns with global candidate coverage for real-world, multi-relational recommendation tasks.

Abstract

Recommendation systems predominantly utilize two-tower architectures, which evaluate user-item rankings through the inner product of their respective embeddings. However, one key limitation of two-tower models is that they learn a pair-agnostic representation of users and items. In contrast, pair-wise representations either scale poorly due to their quadratic complexity or are too restrictive on the candidate pairs to rank. To address these issues, we introduce Context-based Graph Neural Networks (ContextGNNs), a novel deep learning architecture for link prediction in recommendation systems. The method employs a pair-wise representation technique for familiar items situated within a user's local subgraph, while leveraging two-tower representations to facilitate the recommendation of exploratory items. A final network then predicts how to fuse both pair-wise and two-tower recommendations into a single ranking of items. We demonstrate that ContextGNN is able to adapt to different data characteristics and outperforms existing methods, both traditional and GNN-based, on a diverse set of practical recommendation tasks, improving performance by 20% on average.

Figures (1)

• Figure 1: Overview of Context-based Graph Neural Networks.ContextGNN utilizes a bidirectional $\mathrm{GNN}_{\bm{\theta}}$ to learn user $\bm{h}^{(2)}_v$ and user-specific item representations $\bm{h}^{(2)}_w$ within a user's local subgraph. Its message passing scheme is enhanced by additionally propagating shallow item embeddings $\bm{w}_w$ and seed user $\textsc{Identicator}_{\bm{\theta}}$ representations. Afterwards, item scores are produced depending on whether an item situates within a user's subgraph. A user-specific fusion score is learned via an $\mathrm{MLP}_{\bm{\theta}}$ to produce the final ranking by offsetting the contributions of local rankings.