A deeper graph neural network for recommender systems

Abstract

Interaction data in recommender systems are usually represented by a bipartite user–item graph whose edges represent interaction behavior between users and items. The data sparsity problem, which is common in recommender systems, is the result of insufficient interaction data in the link prediction on graphs. The data sparsity problem can be alleviated by extracting more interaction behavior from the bipartite graph, however, stacking multiple layers will lead to over-smoothing, in which case, all nodes will converge to the same value. To address this issue, we propose a deeper graph neural network in this paper that can predict links on a bipartite user–item graph using information propagation. An attention mechanism is introduced to our method to address the problem that variable size inputs for each node on a bipartite graph. Our experimental results demonstrate that our proposed method outperforms five baselines, suggesting that the interactions extracted help to alleviate the data sparsity problem and improve recommendation accuracy.

Introduction

Recommender systems have become increasingly important in recent years due to the problem of information overload. Recommender systems allow individuals to acquire information more effectively by filtering information. Over the years, collaborative filtering has become the most successful and widely used recommendation technique [1], [2]. The core assumption is that users who have expressed similar interests in the past will share similar interests in the future. Popularized by the Netflix Prize, Matrix Factorization (MF) has become the de facto approach to collaborative filtering-based recommendation. Much research effort has been devoted to enhancing the MF method, one of the most powerful collaborative filtering techniques, such as integrating it with neighbor-based models [3], extending it to factorization machines for the generic modeling of features [4], and optimizing it with Bayesian personalized ranking objective to adapt it to implicit feedback recommendation [5]. However, despite these efforts, sparsity is still one of the most challenging issues facing us today.

Recent years have seen a surge in research on graph neural networks, leading to substantial improvements in the performance of tasks with graph-structured data, which is fundamental for recommendation applications. One of the most prominent approaches is Graph Convolutional Networks (GCNs) [6]. The core idea behind GCNs is finding a way to iteratively aggregate feature information about graph structure and the structure of the node’s local graph neighborhood into a machine learning model. The goal is to learn a mapping that embeds nodes as points in a low-dimensional vector space $R^d$. The primary contribution of representation learning approaches is that of finding a way to represent, or encode graph structure which geometric relationships in the embedding space reflect the structure of the original graph. However, GCNs need to operate on the Laplacian eigenbasis which leads to a huge time consumption on large graphs .

In this paper, we view the recommendation task as a link prediction problem on a bipartite graph: the interaction data in collaborative filtering-based methods can be represented by a bipartite graph between user and item nodes as shown in 1. The MF approaches can then be considered as learning a mapping from users/items to a low-dimensional vector, where the interaction information is contained in the vectors.

We propose a general framework named GCF, short for Graph neural network-based Collaborative Filtering, which builds on recent progress in graph neural networks. The framework contains a larger receptive field with iterative information propagation enabling our method to access more information in the process of making decisions. The concept of receptive field on graph neural networks will be given in the following section.

We present an attention-based message-passing method to carry out the information propagation process. In the recommendation task, the variable size input for each layer is a challenge, because the number of neighbors for each node is different. To solve this problem, we assign different weights for neighbors.

Our contributions are as follows:

• We present a general framework GCF to model the latent features of users and items. We also show that MF is a special case of GCF when the number of hidden layers is no more than one.

• We propose an attention-based message-passing method to solve the variable size input problem.

• We perform extensive experiments on three real-world datasets to demonstrate the performance of our GCF method. The results show that the proposed method outperforms the state-of-the-art methods in terms of HR@k and NDCG@k.

The rest of this paper is organized as follows. In Section 2, we formalize the problem and review some of the relevant methods. Section 3 presents our approach in detail. The experiments and results analysis are demonstrated in Section 4. Lastly, conclusions and future work are discussed in Section 5.