Neural Graph Collaborative Filtering

Reviewer: Jolanta MizeraPietraszko

Typically, collaborative filtering (CF) is simply a nearest neighbor (NN) algorithm used either in its original form or in machine learning (ML), especially in supervised learning, to predict user preferences in recommender systems. Here, neural graph collaborative filtering (NGCF) aims to resolve a critical issue of mapping from preexisting features: "the collaborative signal is latent in user-item interactions." The proposed bipartite graph structure integrates this interaction into the embedding process. Some approaches to CF, such as embedding and modeling user-item interactions, suggest that the proposed method is an extension of existing solutions. For example, matrix factorization is replaced with an encoding of the CF signals onto graph representations, named here as high-order connectivity; the presented definition is supported with an example. As neural networks are applicable almost everywhere, the authors "design a neural network method to propagate embeddings recursively on the graph." Following this concept, the recommendation is closely related to the behavioral patterns of users interacting with the same items-the longer the path (more layers), the stronger and more reliable the recommendation. Having the message constructed based on the encoding function, the messages are aggregated and the high-order propagation is computed in matrices whose inner products generate user preferences as to which items are better. The objective function annotates higher prediction values to observed (rather than unobserved) user-item interactions. To avoid overfitting, both message dropout and node dropout are adopted. Experiments conducted on three datasets indicate that the approach allows for a better understanding of user behavior in recommender systems. In my opinion, however, the method does not outperform existing solutions because the assumption relies on the same user-item interaction as in other methods. Perhaps more attributes would boost it. Furthermore, while the method may be interesting, the paper contains some typographical errors (for example, "an vector"); thus, I would rather not recommend it.