Graph Convolutional Embeddings for Recommender Systems

Paula Gómez Duran, Alexandros Karatzoglou, Jordi Vitrià, Xin Xin, Ioannis Arapakis

Abstract

Modern recommender systems (RS) work by processing a number of signals that can be inferred from large sets of user-item interaction data. The main signal to analyze stems from the raw matrix that represents interactions. However, we can increase the performance of RS by considering other kinds of signals like the context of interactions, which could be, for example, the time or date of the interaction, the user location, or sequential data corresponding to the historical interactions of the user with the system. These complex, context-based interaction signals are characterized by a rich relational structure that can be represented by a multi-partite graph. Graph Convolutional Networks (GCNs) have been used successfully in collaborative filtering with simple user-item interaction data. In this work, we generalize the use of GCNs for N-partite graphs by considering N multiple context dimensions and propose a simple way for their seamless integration in modern deep learning RS architectures. More specifically, we define a graph convolutional embedding layer for N-partite graphs that processes user-item-context interactions, and constructs node embeddings by leveraging their relational structure. Experiments on several datasets from recommender systems to drug re-purposing show the benefits of the introduced GCN embedding layer by measuring the performance of different context-enriched tasks.

Introduction

Most modern recommender systems rely on "context"—meta-data like the time of day, a user's location, or the device being used—to improve accuracy. However, traditional models like Factorization Machines (FM) or Neural Collaborative Filtering (NCF) usually treat these context variables as independent features.

While Graph Convolutional Networks (GCNs) have been used to capture the relational structure between users and items, they are typically limited to bipartite graphs. Real-world data is more complex; an interaction is often a multi-dimensional tuple (e.g., User A bought Item B at Location C on Monday). This paper addresses the gap in effectively encoding the topological structure of these $N$-partite interactions into the latent representations (embeddings) used by deep learning models.

Methodology: The Graph Convolutional Embedding (GCE) Layer

The authors propose the Graph Convolutional Embedding (GCE) layer, a plug-and-play component that generalizes GCNs to handle multi-dimensional context. Instead of using a standard lookup table for embeddings, the GCE layer constructs an $N$-partite graph where nodes represent users, items, and various context dimensions (e.g., "Monday," "London," "iPhone"). An edge exists between these nodes if they co-occurred in an interaction.

The proposed architecture integrates an N-partite graph representation into a deep learning pipeline via the Graph Convolutional Embedding (GCE) layer.

Key technical highlights include:

  • Message Passing: The GCE layer uses a modified graph convolution rule. When updating a user node's representation, it aggregates "messages" from the specific context nodes associated with those interactions, rather than just the items.
  • Seamless Integration: The GCE layer is fully differentiable, allowing it to replace the initial embedding layer in existing architectures like Matrix Factorization (MF), FM, or NCF.
  • Relational Inductive Bias: By utilizing the graph structure, the model assumes that entities connected in the graph should have similar representations from the start of training.

Experimental Setup

The approach was evaluated on four distinct benchmarks:

  1. MovieLens 100k: Movie recommendations where the "previous movie clicked" serves as context.
  2. LastFM-1K: Music history where the "previous song" is the context.
  3. User-Book Interactions: A large-scale reading history dataset.
  4. Drug-Disease Interactions: A drug re-purposing task where protein intersections serve as the context.

Results and Discussion

The GCE variants consistently outperformed their standard counterparts across all metrics (Hit Ratio and NDCG). In the MovieLens 100k dataset, the GCE-enhanced Matrix Factorization (MF-GCE) improved NDCG@10 by nearly 67% compared to standard MF.

Performance metrics (NDCG@10 and NDCG@20) across training epochs demonstrate the rapid convergence and superior accuracy of GCE-enhanced models.

Key Insights

  • Relational Bias as a Head Start: GCE-equipped models perform significantly better than baselines after just one epoch of training. This suggests the graph structure captures massive interaction statistics before heavy neural learning begins.
  • The "Equalizer" Effect: GCE provides the biggest boost to simpler models (like MF), helping them close the performance gap with more complex, non-linear models like NCF.
  • Solving the Long Tail: GCE is particularly effective for sparse entities (long-tail items/users). By propagating information across the $N$-partite graph, the model "borrows" knowledge from related nodes to create better embeddings.

Conclusion

The GCE layer serves as a versatile building block for context-aware recommendation. While it shows significant gains, its effectiveness depends on a consistent graph structure; extremely fragmented datasets may see diminished returns. Future work will explore handling dynamic or continuous contextual signals within this graph framework.

Summary updated

Abstract

Modern recommender systems (RS) work by processing a number of signals that can be inferred from large sets of user-item interaction data. The main signal to analyze stems from the raw matrix that represents interactions. However, we can increase the performance of RS by considering other kinds of signals like the context of interactions, which could be, for example, the time or date of the interaction, the user location, or sequential data corresponding to the historical interactions of the user with the system. These complex, context-based interaction signals are characterized by a rich relational structure that can be represented by a multi-partite graph. Graph Convolutional Networks (GCNs) have been used successfully in collaborative filtering with simple user-item interaction data. In this work, we generalize the use of GCNs for N-partite graphs by considering N multiple context dimensions and propose a simple way for their seamless integration in modern deep learning RS architectures. More specifically, we define a graph convolutional embedding layer for N-partite graphs that processes user-item-context interactions, and constructs node embeddings by leveraging their relational structure. Experiments on several datasets from recommender systems to drug re-purposing show the benefits of the introduced GCN embedding layer by measuring the performance of different context-enriched tasks.
BibTeX
@inproceedings{duran2021graph,
  author = {Paula Gómez Duran and Alexandros Karatzoglou and Jordi Vitrià and Xin Xin and Ioannis Arapakis},
  title = {{Graph Convolutional Embeddings for Recommender Systems}},
  booktitle = {IEEE Access},
  year = {2021},
  url = {https://openreview.net/forum?id=U4CzqUzmXI},
}
Figure