[Paper Review] Inductive Graph Pattern Learning for Recommender Systems Based on a Graph Neural Network.

Muhan Zhang, Yixin Chen|arXiv (Cornell University)|Apr 26, 2019

Recommender Systems and Techniques26 references6 citations

TL;DR

This paper proposes Inductive Graph-based Matrix Completion (IGMC), an inductive matrix completion model that learns user and item representations via graph neural networks (GNNs) trained solely on 1-hop subgraphs around (user, item) rating pairs, achieving state-of-the-art performance without side information and enabling generalization to unseen users/items and transfer to new recommendation tasks.

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.

Motivation & Objective

• To develop an inductive matrix completion model that generalizes to unseen users and items without relying on side information.
• To investigate whether local graph patterns around (user, item) pairs can serve as effective predictors of ratings in the absence of side information.
• To evaluate the feasibility of training inductive models for recommendation systems under extreme data scarcity, where only the rating matrix is available.
• To assess the potential of GNNs to capture relevant user-item interaction patterns using only structural information from the rating matrix.
• To explore whether long-range dependencies are necessary for effective recommendation modeling in collaborative filtering settings.

Proposed method

• Constructs 1-hop subgraphs centered on each (user, item) pair from the rating matrix, capturing immediate neighborhood interactions.
• Trains a graph neural network (GNN) to map each 1-hop subgraph to its corresponding rating, using message passing to aggregate neighborhood information.
• Learns user and item embeddings inductively by processing subgraphs during inference, enabling predictions for previously unseen users or items.
• Uses a GNN architecture that processes the subgraph structure and node features (e.g., user/item IDs) to predict the rating of the central (user, item) pair.
• Employs a contrastive learning objective to improve representation quality by encouraging similar subgraphs to have similar embeddings.
• Enables transfer learning by fine-tuning the GNN on a new dataset (e.g., Douban) using only the pre-trained GNN head and new interaction data.

Experimental results

Research questions

• RQ1 Can an inductive matrix completion model achieve competitive performance without using any side information?
• RQ2 Are local 1-hop subgraphs around (user, item) pairs sufficient to predict ratings effectively?
• RQ3 Can a GNN-based model trained on one recommendation dataset generalize to a different dataset without retraining?
• RQ4 Is modeling long-range dependencies necessary for effective collaborative filtering in matrix completion?
• RQ5 How does the inductive capability of the model compare to transductive baselines in terms of zero-shot prediction on unseen users/items?

Key findings

• IGMC achieves performance comparable to or better than state-of-the-art transductive matrix completion methods, even without side information.
• The model generalizes effectively to unseen users and items during inference, demonstrating true inductive bias.
• A model pre-trained on the MovieLens dataset achieves surprisingly strong performance on the Douban movie rating dataset without any fine-tuning, indicating strong transferability.
• Local 1-hop subgraphs alone are sufficient to predict ratings with high accuracy, suggesting that long-range dependencies are not essential for effective recommendation.
• The absence of side information does not hinder performance, as IGMC outperforms or matches transductive baselines in both inductive and transfer learning settings.
• The GNN-based approach successfully learns meaningful user and item representations directly from the rating matrix structure, enabling inductive generalization.