PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Neural Networks in Recommender Systems: A Survey

Shiwen Wu, Fei Sun, Wentao Zhang, Xu Xie, Bin Cui

arXiv:2011.02260v4cs.IRcs.LG

TL;DR

Recommender systems must learn effective user and item representations from interactions and side information amid information overload. This survey classifies GNN-based recommendation models, reviews their challenges and solutions across data and tasks, and discusses future directions. It concludes that GNN-based recommendation research spans diverse model strategies, including graph construction, neighbor aggregation, and information updating, while highlighting unresolved concerns such as uncertainty, dynamics, unreliable relationships, and robustness.

  • Problem

    Recommender systems need effective user and item representations from interactions and side information, while recommendation data have diverse structures and changing or unreliable relationships.

  • Method

    The survey develops a taxonomy of GNN-based recommendation models, reviews representative methods and their strategies, and synthesizes challenges across recommendation tasks and information types.

  • Results

    The survey organizes existing works around graph construction, neighbor aggregation, and information updating, and reviews how representative models address recommendation challenges.

  • Takeaways & Limitations

    The survey provides a unified understanding of recent GNN-based recommender-system research and identifies future directions for the field.

  • Takeaways & Limitations

    Current recommendation studies pay little attention to dynamic graphs, and social relationships may be unreliable or incomplete.

Abstract

from arXiv · show

With the explosive growth of online information, recommender systems play a key role to alleviate such information overload. Due to the important application value of recommender systems, there have always been emerging works in this field. In recommender systems, the main challenge is to learn the effective user/item representations from their interactions and side information (if any). Recently, graph neural network (GNN) techniques have been widely utilized in recommender systems since most of the information in recommender systems essentially has graph structure and GNN has superiority in graph representation learning. This article aims to provide a comprehensive review of recent research efforts on GNN-based recommender systems. Specifically, we provide a taxonomy of GNN-based recommendation models according to the types of information used and recommendation tasks. Moreover, we systematically analyze the challenges of applying GNN on different types of data and discuss how existing works in this field address these challenges. Furthermore, we state new perspectives pertaining to the development of this field. We collect the representative papers along with their open-source implementations in https://github.com/wusw14/GNN-in-RS.

1 INTRODUCTION

Recommender systems address information overload by modeling user preferences from historical interactions, while GNNs provide a graph-based way to represent recommendation data and collaborative signals. This survey organizes GNN-based recommendation research, reviews representative methods and challenges, and identifies future directions.

  • Motivation: Recommender systems help users discover relevant items from large information spaces by modeling preferences from historical interactions.Applications include product, video, music, news, and restaurant recommendations.
  • Why GNNs: Recommendation data naturally form graphs, including user-item bipartite graphs and graphs of item transitions in user behavior sequences.GNNs can therefore provide a unified framework for representing different recommendation information.
  • Why GNNs: GNNs explicitly encode collaborative signals such as graph topology to improve user and item representations.This differs from traditional methods that use interactions mainly as supervised signals.
  • Survey scope: The survey classifies GNN-based recommendation models into five categories based on information type and recommendation task.The categories are user-item collaborative filtering, sequential, social, knowledge graph-based, and other recommendation tasks.
  • Survey scope: For each category, the survey explains key issues, presents representative models, and describes how those models address the issues.The review covers more than 100 shortlisted studies.
  • Future directions: The survey discusses limitations of current methods and proposes nine potential directions for future research.It also summarizes benchmark datasets, evaluation metrics, and real-world applications.

2 BACKGROUNDS AND CATEGORIZATION

This section introduces recommender-system and GNN fundamentals, then categorizes GNN-based recommendation models by information type and task. It explains why graph structure and neighbor aggregation make GNNs applicable across recommendation settings.

  • Recommender Systems: Recommender systems infer users’ preferences from interactions or static features to recommend potentially interesting items.The task estimates a user’s preference for any item using learned user representations and a score function.
  • Recommender Systems: Collaborative filtering jointly learns user and item representations from user-item interaction pairs.Sequential recommendation additionally uses chronological behavior patterns, while session-based recommendation is treated as a subtype under the survey’s broader terminology.
  • Graph Neural Network Techniques: Graphs may be directed or undirected, homogeneous or heterogeneous, and may include hyperedges joining multiple vertices.These graph types provide the structural vocabulary used to represent recommendation data and GNN inputs.
  • Graph Neural Network Techniques: GNNs iteratively aggregate neighbor features and integrate them with each central node’s representation through stacked propagation layers.Aggregation may weight neighbors equally or use attention, while updates can use GRUs, concatenation with nonlinear transformation, or summation.
  • Graph Neural Network Techniques: Typical recommendation-oriented GNNs include GCN, GraphSAGE, GAT, GGNN, and HGNN, which differ in aggregation, attention, recurrence, sampling, or hypergraph modeling.GCN aggregates graph neighbors, GraphSAGE samples fixed-size neighborhoods, GAT differentiates neighbor contributions with attention, GGNN uses GRUs, and HGNN models high-order correlations.
  • Categories of Graph Neural Network-Based Recommendation: The survey classifies GNN-based recommendation models into collaborative filtering, sequential, social, knowledge-graph, and other recommendation tasks.The classification reflects how information type determines graph structure and influences GNN architecture choices such as aggregation, updates, and depth.

3 USER-ITEM COLLABORATIVE FILTERING

GNN-based user-item collaborative filtering models user and item representations by propagating information through interaction graphs. The survey organizes the main design issues around graph construction, neighbor aggregation, information updates, and final node representations.

  • Overview: GNN-based collaborative filtering exploits high-order connectivity in user-item interactions to enhance user and item representations through information diffusion.The pipeline uses interacted items to enhance user representations and users connected to items to enrich item representations.
  • Graph Construction: Graph construction balances representation quality and efficiency by enriching sparse graphs with edges or nodes while sampling neighborhoods on large graphs.Sampling trades off effectiveness against computational efficiency.
  • Neighbor Aggregation: Neighbor aggregation should use attentive weights for heterogeneous neighbors, whereas equal weighting or degree normalization is preferable when neighbors are less heterogeneous.Explicitly modeling central-neighbor affinity or interactions among neighbors may help but requires validation on more datasets.
  • Information Update: Information updates generally retain both the central node representation and aggregated neighbor information, while simplified GCN updates can improve performance and computational efficiency.Recent methods remove transformation and non-linearity operations while retaining or improving performance.
  • Final Node Representation: Final user/item representations preferably combine representations from all layers rather than using only the last layer.Weighted pooling differentiates layer contributions, while concatenation preserves information from all layers.

4 SEQUENTIAL RECOMMENDATION

Sequential recommendation uses GNNs to model item-transition patterns after transforming user behavior sequences into graphs. Key design choices concern graph construction, information propagation, and integrating sequential preferences.

  • Overview: Sequential recommendation models users’ next preferences by representing item transitions in their recent behavior sequences as graph-structured data.GNNs are applied after sequential behaviors are transformed into sequence graphs.
  • Graph Construction: Graph construction must address whether to build independent sequence subgraphs, connect nonconsecutive items, or enrich sequences with additional behavioral data.Additional sequences may include other behavior types, the same user’s history, or sequences from the broader dataset.
  • Graph Construction: Short sequences can produce sparse graphs, motivating added transitions, modified current-sequence structures, or hypergraphs that capture higher-order and cross-session relations.The preprocessed Yoochoose1/41 dataset has an average sequence length of 5.71.
  • Information Propagation: Propagation methods aggregate preceding and following items with mean pooling, GRU-based updates, attention, or hypergraph mechanisms, while some methods preserve neighborhood order.Permutation-invariant aggregation can lose item-order information; more complex propagation may improve performance at higher computational cost.
  • Sequential Preference: Attention commonly integrates item representations for sequential preference, positional embeddings provide additional order information, and the value of RNN structures remains unresolved.The survey reports no consensus on the best graph-construction or propagation method and calls for further investigation of RNNs.

5 SOCIAL RECOMMENDATION

Social recommendation uses users’ social relationships to enhance preference modeling, while GNNs capture influence beyond immediate friends. The main design issues are weighting friends and integrating social and interaction information.

  • Overview: GNNs model recursive social diffusion because users may be influenced by friends’ friends, beyond the first-order relationships represented in earlier approaches.Overlooking higher-order influence can lead to suboptimal recommendation performance.
  • Core Issues: Social recommendation must determine whether friends have equal influence and how to combine social relationships with user-item interactions.These are the section’s two central modeling issues.
  • Influence of Friends: Attention-based influence modeling improves overall performance compared with mean pooling, supporting differentiated influence weights for different friends.The cited comparison is reported for a model combining recurrent behavior modeling with graph attention.
  • Influence of Friends: Social relationships may be unreliable because explicit connections can lack influence and observed relationships can be incomplete.Methods such as ESRP and DiffNetLG filter or infer relationships before using them for recommendation.
  • Preference Integration: Two integration strategies learn representations from social and interaction graphs separately before combining them, or unify both graphs for joint propagation.Separate graphs allow different diffusion depths, whereas unified graphs simulate social and interest diffusion together.

6 KNOWLEDGE GRAPH BASED RECOMMENDATION

Knowledge-graph recommendation combines user-item interactions with semantic item relations to improve preference estimation. GNN-based designs must construct suitable graphs and aggregate information in a relation-aware way.

  • Overview: Knowledge graphs provide semantic relations among items that can reveal connections and improve item representations for recommendation.They are used alongside user-item interaction information to estimate users’ item preferences.
  • Core Issues: The main challenges are integrating collaborative and semantic information during graph construction and aggregating linked entities across multiple relation types.These challenges determine whether user nodes are explicit graph entities or implicit signals for relation weighting.
  • Graph Construction: Existing methods either incorporate users into unified or user-item-specific knowledge-graph subgraphs, or use users implicitly to weight relations.User-item subgraphs focus on more related entities and relations but require more computation and depend on graph construction quality.
  • Relation-aware Aggregation: Relation-aware propagation considers linked entities, their connecting relations, and sometimes user roles, commonly through variants of graph attention networks.When users are explicit entities, their preferences diffuse through interacted items and related attributes; otherwise, user representations weight relations directly.

7 OTHER TASKS

GNNs have been applied beyond the survey’s four main recommendation categories, including POI, group, bundle, CTR, and multimedia recommendation. These tasks use graphs suited to geographic, group, feature, collaborative, or multimodal structure.

  • POI Recommendation: POI recommendation combines user-POI interactions, check-in sequence graphs, and geographic graphs to model geographical influence and transition patterns.Geographic edges connect POIs within a distance threshold, with edge weights depending on geographic information.
  • Group Recommendation: Group recommendation models user-item, user-group, and group-item relationships, with groups serving as bridges between users and items.The group may be treated as part of the graph or handled separately.
  • Bundle Recommendation: Bundle recommendation represents users, bundles, and bundle items through unified or hierarchical propagation, including item-item interactions within bundles.The central challenge is obtaining a useful bundle representation.
  • CTR Prediction: CTR models use GNNs to capture high-order feature interactions, while later designs also incorporate user-item collaborative signals and similarity-based relations.Fi-GNN constructs a feature graph; DG-ENN adds attribute and collaborative graphs to address omitted collaborative signals and interaction sparsity.
  • Multimedia Recommendation: Multimedia recommendation applies GNNs to modality-specific user-item graphs to learn representations from interactions with multimodal content.MMGCN propagates information separately for each modality graph.

8 DATASETS, EVALUATION METRICS AND APPLICATIONS

This section surveys datasets, evaluation metrics, and real-world applications used in GNN-based recommendation research. It highlights commonly adopted benchmarks across recommendation tasks and explains core ranking metrics.

  • Datasets: The survey catalogs datasets for user-item, social, knowledge-graph, sequential, session-based, and POI recommendation tasks.Examples include MovieLens, Epinions, Ciao, Book-Crossing, Amazon, Yelp, Gowalla, Yoochoose, Diginetica, RetailRocket, and LastFM.
  • Datasets: MovieLens datasets provide rating pairs, timestamps, movie attributes and tags, and user demographic features for benchmark evaluation.MovieLens-100K, MovieLens-1M, and MovieLens-20M differ in scale, with ratings ranging from 1 to 5.
  • Datasets: Amazon datasets combine reviews, product metadata, and also-viewed or also-bought links for collaborative filtering and sequential recommendation.Category-specific subsets include Amazon-Books, Amazon-Instant Video, and Amazon-Electronics.
  • Evaluation Metrics: Precision@K, Recall@K, and F1@K evaluate top-K recommendation accuracy, while NDCG accounts for the ranking positions of correctly recommended items.AUC evaluates whether clicked items are ranked above non-clicked items, and MAP measures average precision over users.
  • Applications: GNN-based recommendation models are applied in industry, including product or advertisement recommendation on e-commerce platforms.IntentGC uses user-item preferences and heterogeneous knowledge-graph relationships and is deployed at Alibaba.

9 FUTURE RESEARCH DIRECTIONS AND OPEN ISSUES

The survey identifies future directions involving richer user-interest representations and scalable GNNs. It emphasizes modeling uncertainty and multiple interests while addressing industrial graph size and training costs.

  • Diverse and Uncertain Representation: A single vector struggles to capture users’ diverse and uncertain interests, motivating representations that model multiple interests.The survey identifies this as an open direction for GNN-based recommendation.
  • Diverse and Uncertain Representation: Multi-vector approaches include disentangled representations and capsule networks, with DGCF modeling multi-aspect representations using orthogonal constraints.DGCF also iteratively updates adjacent relationships for each aspect.
  • Diverse and Uncertain Representation: Density-based representations such as Gaussian embeddings can encode uncertainty and asymmetric relationships more naturally than several point-based similarities.Gaussian embedding has been used to model uncertain preferences, but remains insufficiently studied in GNN-based recommendation.
  • Scalability: Industrial recommendation graphs can contain billions of nodes and edges and millions of features per node, making traditional GNNs costly in memory and training time.The survey describes graph sampling and scalable architectural changes as the two mainstream responses.
  • Scalability: Sampling improves scalability but loses information, leaving effective sampling strategies that balance effectiveness and scalability as an open problem.GraphSAGE samples fixed-size neighborhoods, while PinSage uses random walks.
  • Scalability: Decoupling nonlinearities and weight-matrix operations improves scalability through one-time precomputation but limits aggregator and updater flexibility.These models therefore trade architectural flexibility for reduced training communication cost.

9.3 Dynamic Graphs in Recommendation

Recommendation graphs are dynamic because users, items, and their relationships change over time. The survey highlights adaptive propagation and incremental learning as underexplored responses to changing graph structure.

  • Dynamic Graphs: Changing users, items, and relationships turn real-world recommendation data into dynamic rather than static graphs.Systems must update iteratively with newly arriving information to maintain current recommendations.
  • Dynamic Graphs: GraphSAIL addresses incremental GNN learning by constraining embedding similarity to balance updates with preservation of prior representations.Its setting focuses on changing interactions, namely edges between nodes.
  • Adaptive Propagation: Recommendation node degrees follow a long-tail distribution, so active and cold users or items may require different propagation depths.Using the same propagation step for every node may produce suboptimal receptive fields.
  • Adaptive Propagation: Adaptive propagation remains an emerging direction because only a few studies decide propagation steps separately for each node.The goal is to obtain a reasonable receptive field for nodes with different degrees.

9.5 Self-supervised Learning

The survey presents self-supervised learning as a promising way to address sparsity in GNN-based recommendation, while also identifying robustness and privacy as practical challenges.

  • Self-supervised Learning: Self-supervised learning can improve data utilization and help alleviate sparsity in recommender systems.Few GNN-based recommendation studies had adopted it at the time of the survey.
  • Self-supervised Learning: COTREC maximizes agreement between the last-clicked item and predicted-item samples alongside the session representation.This constructs a contrastive learning task for session-based recommendation.
  • Self-supervised Learning: DHCN maximizes mutual information between session representations learned from a session-to-session graph and an item-session hypergraph.The survey identifies designing an effective task-aligned supervised signal as the key challenge.
  • Robustness: Small input perturbations can sharply reduce GNN performance, while unreliable clicks, incomplete social relationships, and injected fake data threaten recommendation robustness.The survey calls for stable recommendations under shilling attacks.
  • Privacy: Privacy-preserving GNN recommendation uses federated learning and local user-item graphs, but local differential privacy may reduce model accuracy through added gradient noise.FedGNN combines pseudo-interacted items with graph expansion to protect items and exploit higher-order interactions.

9.8 Fairness in GNN-based Recommender System

Fairness and explainability are important open concerns for GNN-based recommender systems. Existing work addresses recommendation bias and studies both instance-level and model-level explanations.

  • Fairness: Fairness research seeks comparable recommendation performance across demographic groups and equal overall exposure probabilities for items.The passage identifies discriminatory decisions by GNNs as a societal concern and calls for further study of fairness alongside comparable performance.
  • Fairness: NISER uses representation normalization to address popularity bias, while FairGNN uses adversarial learning with limited sensitive information.These are representative strategies for mitigating bias in GNN-based recommender systems.
  • Explainability: Explainable recommendations can clarify why items are suggested for users and help practitioners understand models for further improvements.The passage also notes that explanations may make recommendations more persuasive.
  • Explainability: GNN explainability methods divide into instance-level explanations of individual predictions and model-level interpretations of deep graph models.Instance-level methods identify important input features, whereas model-level methods provide generic understanding of model behavior.

10 CONCLUSION

The survey reviews recent GNN-based recommender-system research and organizes it into a classification scheme. It explains representative strategies, their advantages and limitations, and proposes directions for future research.

  • Conclusion: The survey provides a comprehensive review of recent research on GNN-based recommender systems.It aims to give readers a general understanding of recent progress in the field.
  • Conclusion: Its classification scheme organizes existing works and clarifies issues, strategies, advantages, and limitations for each category.The survey uses representative models to detail how different categories address their main issues.
  • Conclusion: The survey suggests several promising directions for future research in GNN-based recommender systems.It presents these directions as guidance for future developments in the field.
Loading 2011.02260v4…