Inductive Graph Pattern Learning for Recommender Systems Based on a Graph Neural Network

Muhan Zhang, Yixin Chen

Abstract

We propose an inductive matrix completion model without using side information. By factorizing the (rating) matrix into the product of low-dimensional latent embeddings of rows (users) and columns (items), a majority of existing matrix completion methods are transductive, since the learned embeddings cannot generalize to unseen rows/columns or to new matrices. To make matrix completion inductive, most previous works use content (side information), such as user's age or movie's genre, to make predictions. However, high-quality content is not always available, and can be hard to extract. Under the extreme setting where not any side information is available other than the matrix to complete, can we still learn an inductive matrix completion model? In this paper, we propose an Inductive Graph-based Matrix Completion (IGMC) model to address this problem. IGMC trains a graph neural network (GNN) based purely on 1-hop subgraphs around (user, item) pairs generated from the rating matrix and maps these subgraphs to their corresponding ratings. It achieves highly competitive performance with state-of-the-art transductive baselines. In addition, IGMC is inductive -- it can generalize to users/items unseen during the training (given that their interactions exist), and can even transfer to new tasks.

Introduction

Most recommender systems rely on Matrix Factorization, which is inherently transductive. This means the model learns specific "embeddings" for every user and item in the training set. If a new user joins or a new movie is added, the entire model often needs to be retrained because it has no latent vector for these new entities. To solve this, researchers developed Inductive Matrix Completion (IMC). However, traditional IMC relies heavily on side information (content)—such as a user's age or a movie's genre—to make predictions for new entries. The problem is that high-quality side information is often unavailable, noisy, or difficult to extract. IGMC addresses the bold question: Can we build an inductive recommender system using only the rating matrix itself?

The IGMC Framework

IGMC treats matrix completion as a link-prediction task on a bipartite graph. The framework follows a multi-step process to learn structural patterns without relying on node identities.

Overview of the IGMC pipeline: from rating matrix to subgraph extraction and pattern learning.
  1. Enclosing Subgraph Extraction: For any user-item pair $(u, v)$ whose rating needs to be predicted, the model extracts an $h$-hop local subgraph containing the target user, the target item, and their immediate neighbors.
  2. Node Labeling: To make the GNN aware of the target nodes, IGMC applies a unique labeling scheme. The target user is labeled 0, the target item is labeled 1, and other nodes are labeled based on their distance from the targets. This allows the model to learn structural patterns (e.g., "users who liked what you liked also liked this") without needing to know specific user IDs.
  3. Relational GNN (R-GCN): Since ratings are discrete (e.g., 1–5 stars), the model treats each rating value as a different "edge type," using a Relational Graph Convolutional Network to aggregate information.
  4. Adjacent Rating Regularization (ARR): This loss term encourages the model's internal parameters for similar ratings (like "4" and "5") to be similar, reflecting the natural ordinal relationship of ratings.
  5. Graph-Level Pooling: Finally, the model concatenates the representations of the target user and item to output a final predicted rating.

Experimental Results

The model was evaluated on five standard collaborative filtering benchmarks, including Flixster, Douban, YahooMusic, and MovieLens datasets.

RMSE comparison of IGMC against inductive and transductive baselines on three benchmark datasets.

As shown in the results, IGMC achieved the lowest Root Mean Square Error (RMSE) on Flixster (0.872), Douban (0.721), and YahooMusic (19.1), outperforming models that had access to side information. On MovieLens-100K, IGMC matched the performance of the best transductive models (RMSE 0.905) while remaining fully inductive.

Key Insights

  • Local is Enough: The paper demonstrates that 1-hop or 2-hop neighborhoods provide enough structural information to predict ratings accurately. Long-range dependencies might not be necessary for modeling recommender systems.
  • Transfer Learning: Because the model learns general interaction patterns, a model trained on MovieLens (Western movies) can perform well on Douban (Chinese movies) without retraining. This suggests that "recommendation logic" is somewhat universal across platforms.
  • Robustness to Sparsity: IGMC significantly outperforms traditional methods when data is extremely sparse, maintaining higher accuracy even with only 0.1% of ratings.

Conclusion

IGMC demonstrates that it is possible to train inductive matrix completion models without using side information while achieving similar or better performances than state-of-the-art transductive methods. While it still requires at least one interaction to form a subgraph (the "extreme cold-start" remains a challenge), it offers a powerful, identity-agnostic approach to recommendation that generalizes across users, items, and datasets.

Summary updated

Abstract

We propose an inductive matrix completion model without using side information. By factorizing the (rating) matrix into the product of low-dimensional latent embeddings of rows (users) and columns (items), a majority of existing matrix completion methods are transductive, since the learned embeddings cannot generalize to unseen rows/columns or to new matrices. To make matrix completion inductive, most previous works use content (side information), such as user's age or movie's genre, to make predictions. However, high-quality content is not always available, and can be hard to extract. Under the extreme setting where not any side information is available other than the matrix to complete, can we still learn an inductive matrix completion model? In this paper, we propose an Inductive Graph-based Matrix Completion (IGMC) model to address this problem. IGMC trains a graph neural network (GNN) based purely on 1-hop subgraphs around (user, item) pairs generated from the rating matrix and maps these subgraphs to their corresponding ratings. It achieves highly competitive performance with state-of-the-art transductive baselines. In addition, IGMC is inductive -- it can generalize to users/items unseen during the training (given that their interactions exist), and can even transfer to new tasks. Our transfer learning experiments show that a model trained out of the MovieLens dataset can be directly used to predict Douban movie ratings with surprisingly good performance. Our work demonstrates that: 1) it is possible to train inductive matrix completion models without using side information while achieving similar or better performances than state-of-the-art transductive methods; 2) local graph patterns around a (user, item) pair are effective predictors of the rating this user gives to the item; and 3) Long-range dependencies might not be necessary for modeling recommender systems.
BibTeX
@inproceedings{zhang2020inductive,
  author = {Muhan Zhang and Yixin Chen},
  title = {{Inductive Graph Pattern Learning for Recommender Systems Based on a Graph Neural Network}},
  booktitle = {International Conference on Learning Representations},
  year = {2020},
  url = {https://openreview.net/forum?id=OOxqSvG7niB},
}
Figure