A Survey of Graph Neural Networks for Social Recommender Systems

Kartik Sharma; Yeon-Chang Lee; Sivagami Nambi; Aditya Salian; Shlok Shah; Sang-Wook Kim; Srijan Kumar

Source-linked AI summary

A Survey of Graph Neural Networks for Social Recommender Systems

Kartik Sharma, Yeon-Chang Lee, Sivagami Nambi, Aditya Salian, Shlok Shah, Sang-Wook Kim, Srijan Kumar

arXiv:2212.04481v3Dec 8, 2022cs.SIcs.IRcs.LG

TL;DR

SocialRS research increasingly uses GNNs to combine user-item interactions with user-user social relations, but the GNN-based literature lacked a dedicated comprehensive survey. This paper systematically reviews the field using PRISMA, identifies 84 papers, and organizes them through taxonomies of inputs and architectures. It also summarizes datasets, metrics, and future research directions.

Problem
Existing surveys focused on traditional SocialRS, selected features or applications, or general graph-based recommendation, without comprehensively covering GNN-based SocialRS.
Method
The survey follows PRISMA to identify relevant literature, manually filters the papers, and reviews methods through taxonomies of inputs and architectures.
Results
The survey identifies 84 GNN-based SocialRS papers and organizes them into input and architecture taxonomies covering input types, representations, GNN encoders, decoders, and loss functions.
Takeaways & Limitations
The survey provides a structured view of GNN-based SocialRS methods, benchmark datasets, evaluation metrics, and future research directions.
Takeaways & Limitations
Existing methods face limitations including narrow graph-augmentation strategies, while the survey identifies scalability as an unresolved challenge for large and rapidly growing graphs.

Abstract

from arXiv · show

Social recommender systems (SocialRS) simultaneously leverage the user-to-item interactions as well as the user-to-user social relations for the task of generating item recommendations to users. Additionally exploiting social relations is clearly effective in understanding users' tastes due to the effects of homophily and social influence. For this reason, SocialRS has increasingly attracted attention. In particular, with the advance of graph neural networks (GNN), many GNN-based SocialRS methods have been developed recently. Therefore, we conduct a comprehensive and systematic review of the literature on GNN-based SocialRS. In this survey, we first identify 84 papers on GNN-based SocialRS after annotating 2151 papers by following the PRISMA framework (preferred reporting items for systematic reviews and meta-analyses). Then, we comprehensively review them in terms of their inputs and architectures to propose a novel taxonomy: (1) input taxonomy includes 5 groups of input type notations and 7 groups of input representation notations; (2) architecture taxonomy includes 8 groups of GNN encoder notations, 2 groups of decoder notations, and 12 groups of loss function notations. We classify the GNN-based SocialRS methods into several categories as per the taxonomy and describe their details. Furthermore, we summarize benchmark datasets and metrics widely used to evaluate the GNN-based SocialRS methods. Finally, we conclude this survey by presenting some future research directions. GitHub repository with the curated list of papers are available at https://github.com/claws-lab/awesome-GNN-social-recsys.

1 INTRODUCTION

Social recommender systems combine user-item interactions with user-user social relations, while GNNs have recently motivated a growing body of methods. This survey addresses the lack of a comprehensive review by organizing the literature around inputs and architectures.

Motivation: SocialRS uses user-item interactions and user-user social relations to recommend items, leveraging homophily and social influence to understand users’ tastes.Social relations can also help mitigate data sparsity and support recommendation across product, music, location, and image domains.
Motivation: MF-based methods struggle to model complex nonlinear relationships, motivating the use of deep-learning methods including GNNs.User-item interactions and user-user relations can naturally be represented as graph data.
Challenges: GNN-based SocialRS must jointly extract knowledge from interaction and social networks while incorporating features such as user, item, knowledge, and group information.Many methods represent these inputs using separate user-item and user-user graphs before fusing network and feature information.
Challenges: GNN-based SocialRS research requires choices about encoder designs and training losses that produce embeddings reflecting user tastes and item characteristics.Existing methods use architectures such as GANNs and losses including MSE, BPR, CE, and auxiliary objectives.
Related Surveys: Earlier surveys emphasized traditional SocialRS, feature information, specific applications, or general graph-based recommendation, leaving GNN-based SocialRS insufficiently covered.The survey positions itself as the first systematic review focused specifically on GNN-based SocialRS.
Contributions: The survey contributes a PRISMA-based review, a taxonomy of inputs and architectures, summaries of publication venues and method trends, and guidance for future research.Its taxonomy covers 5 input-type groups, 7 input-representation groups, 8 encoder groups, 2 decoder groups, and 12 loss-function groups.

2 SURVEY METHODOLOGY

The survey follows PRISMA-guided retrieval and manual annotation to identify relevant GNN-based SocialRS literature. Four expert annotators screened 2,151 records and produced a final set of 84 papers.

Retrieval: 2,151 papers were retrieved from Scopus using a query covering social recommendation, recommender systems, graphs, publication year, and English-language restrictions.The query was run on October 14, 2022.
Annotation: Four expert annotators manually reviewed and filtered the retrieved papers under PRISMA guidelines after agreeing on definitions for the inclusion concepts.The agreed concepts included graph-related and social-recommendation criteria.
Annotation: Each annotator labeled an initial shared batch using Yes, No, or Maybe categories, with an inter-annotator agreement of 0.845.Yes indicated full confidence in relevance, No full confidence in irrelevance, and Maybe some confidence in relevance.
Final Selection: 84 papers remained after independent annotation and consensus review of papers marked Maybe.Papers labeled Yes formed the survey’s final study set.

3 NOTATIONS AND PROBLEM DEFINITION

The paper represents social recommendation with user-item ratings and user-user social relations, then defines rating-prediction and top-N recommendation objectives over unrated items.

Notation: The formulation uses m users, n items, a rating matrix R ∈ R^m×n, a social matrix S ∈ R^m×m, and each user’s rated-item set N_pᵢ.Bold uppercase and lowercase letters denote matrices and vectors, while calligraphic letters denote sets and graphs.
Problem Definition: GNN-based SocialRS methods target rating prediction and/or top-N recommendation given the rating matrix R and social matrix S.The two tasks are defined over items a user has not rated.
Problem Definition: Rating prediction estimates ratings for each user’s unrated items as close as possible to the ground truth.The unrated-item set is I\N_pᵢ.
Problem Definition: Top-N recommendation selects the N unrated items most likely to be preferred by each user.Recommendations are made from the user’s unrated-item set I\N_pᵢ.

4 TAXONOMY OF INPUTS

The survey organizes GNN-based SocialRS inputs by both the data types they use and the graph representations adopted to encode them. It identifies five input-type categories and seven input-representation categories, covering separate, unified, attributed, multiplex, hypergraph, knowledge-graph-enhanced, and decentralized designs.

4.1 Input Types: Five input types are identified: user-item ratings, user-user social relations, attributes, knowledge graphs, and groups.Table 3 categorizes surveyed papers according to these input data types.
4.1 Input Types: User-item ratings record interactions, timestamps, and potentially multiple interaction types such as positive and negative feedback.Timestamps support recommendations at particular times, while some methods distinguish interaction types during prediction.
4.1 Input Types: Social inputs are stored as a user-user adjacency matrix and may represent friendships, co-commenting, following, or other heterogeneous relations.Some methods explicitly model multifaceted user-user relations.
4.1 Input Types: Additional features include user or item attributes, item-item knowledge-graph dependencies, and user groups based on shared interests, hobbies, businesses, or clients.These inputs supplement interaction and social-network information in different method designs.
4.2 Input Representations: Seven input representations are distinguished: U-U/U-I graphs, U-U-I graphs, attributed graphs, multiplex graphs, U-U/U-I/I-I graphs, hypergraphs, and decentralized structures.Table 4 categorizes papers by the graph representation developed from their input data.
4.2 Input Representations: Separate U-U and U-I graphs encode networks independently before aggregation, whereas a U-U-I graph merges social and interaction edges into one graph.Attributed and multiplex variants add node features or multiple relation layers, respectively.
4.2 Input Representations: Item-item graphs, hypergraphs, and decentralized designs extend representation to higher-order relations, knowledge dependencies, or privacy-preserving local storage.Hypergraphs can connect user groups, users with shared items, or users with groups of items; decentralized designs retain local sensitive edges.

5 TAXONOMY OF ARCHITECTURES

GNN-based SocialRS architectures are organized around encoders, decoders, and loss functions that transform social and interaction graphs into recommendation scores. The survey further categorizes encoder designs and their computational trade-offs.

Architecture Components: Architectures contain encoders, decoders, and loss functions: encoders produce user/item embeddings, decoders predict preferences, and losses train embeddings end-to-end.Additional user or item information can enhance embeddings, while dashed flows represent auxiliary inputs or losses.
Encoders: The survey groups encoders into 8 categories: GCN, LightGCN, GANN, HetGNN, GRNN, HyperGNN, GAE, and hyperbolic GNN.Some surveyed methods instead use RNN, MLP, embedding vectors, or no encoder.
Encoders: Most methods represent users with interaction and social embeddings, aggregate them into one user embedding, and obtain item embeddings from user-item graphs.Some methods use one encoder for both embeddings, whereas others use distinct encoders for different node types; attributes and hypergraphs may provide additional representations.
Encoders: GCN encoders aggregate neighbor representations using nonlinear activation and trainable transformations, with final embeddings taken from the last layer or aggregated across layers.Self-connections may also be included during neighborhood aggregation.
Encoders: LightGCN removes nonlinear activation, feature transformation, and self-connection to simplify propagation, while attention mechanisms weight neighbors to emphasize important adjacent nodes.SocialRS methods use concatenation-based or similarity-based attention, with similarity functions such as cosine similarity and dot product.
Encoders: HetGNN models user-item and user-user relations as heterogeneous relationships and uses different transformation matrices according to relation types.One example defines directed user-user, user-item, item-user, and item-item edges.
Decoders: Decoders are grouped into 2 categories: dot-product and multi-layer perceptron (MLP).The dot-product decoder predicts preference from user and item embeddings through their inner product.
Loss Functions: Loss functions comprise 4 primary categories—BPR, MSE, CE, and hinge loss—and 8 auxiliary categories, including social link prediction, self-supervised, adversarial, and knowledge-distillation losses.The survey also lists group-based, path-based, sentiment-aware, and policy-network-based auxiliary losses.

6 EXPERIMENTAL SETUP

The survey reviews benchmark datasets and evaluation practices for GNN-based SocialRS, then compares recommendation accuracy across representative application domains. The datasets span diverse domains and provide varying combinations of user-item interactions and user-user relations.

Benchmark datasets: 17 benchmark datasets across eight application domains are reviewed, with statistics on users, items, ratings, social relations, and dataset usage.The domains are product, location, movie, image, music, bookmark, microblog, and miscellaneous.
Benchmark datasets: Table 11 distinguishes datasets containing both user-item interactions and user-user relations from those missing one relation type.The table caption states that dataset colors encode whether both relation types are available.
Benchmark datasets: The reviewed datasets cover products, locations, movies, images, music, bookmarks, microblogs, and miscellaneous applications, with substantial variation in scale and feedback type.Examples include Epinions, Ciao, Yelp, Gowalla, Foursquare, MovieLens, Flixster, FilmTrust, Flickr, Last.fm, and Delicious.
Benchmark datasets: MovieLens methods construct social relations from user similarities because the original dataset lacks users’ social relations.The cited version contains 487.1K social relations and 1.5M ratings from 138.1K users on 16.9K movies.
Experimental results: Accuracy comparisons use one dataset per domain and include only methods evaluated under the same settings on each dataset.The comparison covers Epinions, Yelp, Flixster, Flickr, Last.fm, Delicious, and Douban.
Experimental results: No evidence indicates that a GNN encoder is optimized for a specific domain; the best performer varies with domain and metric.The survey notes that many Douban methods use HyperGNN, possibly to capture varied motifs from different user behaviors and item types.

7 FUTURE DIRECTIONS

The survey identifies sparsity, trustworthiness, graph heterogeneity, and scalability as open challenges for GNN-based SocialRS. It proposes richer augmentation and self-supervision, trustworthy architectures, heterogeneous modeling, and more scalable designs as future directions.

7.5 Experimental comparison: Table 12 compares recommendation accuracy only among methods with the same settings on each dataset.This comparison criterion frames the survey’s domain-level conclusions about varying best performers.
7.1 Graph Augmentation in GNN-based SocialRS: Sparse user-item interactions and user-user relations motivate using additional supervision signals and graph views beyond the original structure.Existing methods mainly add edges between users or between users and items, leaving broader augmentation techniques underexplored.
7.2 Trustworthy GNN-based SocialRS: Trustworthy GNN-based SocialRS should address robustness, explainability, privacy, and fairness in addition to recommendation accuracy.The survey notes one privacy-oriented federated-learning framework and identifies robustness against targeted attacks as unanswered.
7.3 Heterogeneous GNN-based SocialRS: Only a few SocialRS studies leverage heterogeneous graphs, despite their multi-typed nodes and interactions, heterogeneous attributes, meta-paths, and temporal properties.Designing HetGNN-based SocialRS remains an open direction.
7.4 Efficiency and Scalability: Most real-world graphs are large and rapidly growing, while many GNN-based SocialRS methods remain too complicated to scale efficiently.Scalability attempts simplify model components or use knowledge distillation, but highly scalable architectures remain challenging.

8 CONCLUSIONS

This survey systematically reviews 84 GNN-based SocialRS papers identified from 2,151 papers using PRISMA guidelines. It introduces taxonomies of inputs and architectures and summarizes datasets, metrics, and future research directions.

Conclusions: The survey is the first systematic and comprehensive review of 84 GNN-based SocialRS papers collected following PRISMA guidelines.It addresses the absence of a thorough survey in this area.
Conclusions: Its taxonomy categorizes GNN-based SocialRS methods by their inputs and architectures to organize developments in the field.The survey also summarizes benchmark datasets and evaluation metrics and presents future research directions.