PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Graph Neural Networks for Recommender Systems: Challenges, Methods, and Directions

Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, Yong Li

arXiv:2109.12843v3cs.IR

TL;DR

GNNs have become a major approach to recommender systems, but applying them raises challenges in graph construction, model design, optimization, and efficiency. This survey systematically reviews the field, organizing methods across four recommender-system perspectives and discussing open problems and future directions. It also summarizes representative papers and code repositories to support research in the area.

  • Problem

    Recommender systems need to model high-order connectivity and structural information beyond directly observed interactions, while GNN-based approaches introduce important design and efficiency challenges.

  • Method

    The paper provides a systematic survey of GNN-based recommender systems, covering their history, motivations, challenges, methods, open problems, and future directions through taxonomies of stage, scenario, objective, and application.

  • Results

    The survey comprehensively presents the challenges, representative methods, recent advances, and future directions of GNN-based recommender systems.

  • Takeaways & Limitations

    The taxonomy and repository of representative papers provide a structured resource for understanding and developing GNN-based recommender systems.

  • Takeaways & Limitations

    GNN models have substantially higher computation costs than traditional recommendation methods, so neighbor sampling or graph pruning can improve efficiency while reducing recommendation performance.

Abstract

from arXiv · show

Recommender system is one of the most important information services on today's Internet. Recently, graph neural networks have become the new state-of-the-art approach to recommender systems. In this survey, we conduct a comprehensive review of the literature on graph neural network-based recommender systems. We first introduce the background and the history of the development of both recommender systems and graph neural networks. For recommender systems, in general, there are four aspects for categorizing existing works: stage, scenario, objective, and application. For graph neural networks, the existing methods consist of two categories, spectral models and spatial ones. We then discuss the motivation of applying graph neural networks into recommender systems, mainly consisting of the high-order connectivity, the structural property of data, and the enhanced supervision signal. We then systematically analyze the challenges in graph construction, embedding propagation/aggregation, model optimization, and computation efficiency. Afterward and primarily, we provide a comprehensive overview of a multitude of existing works of graph neural network-based recommender systems, following the taxonomy above. Finally, we raise discussions on the open problems and promising future directions in this area. We summarize the representative papers along with their code repositories in \url{https://github.com/tsinghua-fib-lab/GNN-Recommender-Systems}.

1 INTRODUCTION

The survey positions GNNs as a major approach to recommender systems and organizes the field’s development, challenges, methods, and directions through a systematic taxonomy. It emphasizes high-order connectivity, structural data, and comprehensive coverage of recent GNN-based recommendation research.

  • Development of Recommender Systems: Recommendation research developed from shallow models to neural models and then GNN-based models, with early methods capturing collaborative filtering through interaction similarity or representation learning.Traditional approaches include matrix factorization and factorization machines, but they face challenges involving complex user behaviors.
  • Motivation for GNNs: GNNs aggregate neighborhood embeddings through iterative propagation, allowing nodes to access high-order neighbors rather than only first-order neighbors.This capability helps capture structural information in recommender-system data and has made GNN-based methods state-of-the-art approaches in the area.
  • Research Challenges: Applying GNNs to recommender systems requires addressing graph construction, propagation and aggregation design, model optimization, and computation efficiency.The input must represent elements as nodes and relations as edges, while GNN components need to be adapted to each task.
  • Survey Scope: The survey categorizes recommender-system research by stage, scenario, objective, and application, then reviews representative GNN-based methods using this taxonomy.It also summarizes representative papers together with their code repositories.
  • Related Surveys: Compared with earlier surveys, this work provides an extensive and up-to-date review focused specifically on GNN-based recommenders.Prior surveys covered only small fractions of GNN-based recommendation methods or organized related areas differently.

2.1 Recommender Systems

Recommender systems are organized by stages, scenarios, objectives, and applications, with a standard multi-stage pipeline and diverse task settings. Beyond accuracy, systems address diversity, explainability, and fairness across applications such as e-commerce and point-of-interest recommendation.

  • Stages: Recommender systems are commonly organized into matching, ranking, and re-ranking stages that progressively filter a large item pool.Matching retrieves candidate items efficiently, ranking captures preferences, and re-ranking considers criteria beyond accuracy.
  • Scenarios: Social recommendation incorporates social relations because user behavior can reflect both personal preferences and social factors.Social influence and social homophily are described as mechanisms shaping interactions.
  • Scenarios: The survey distinguishes scenarios including social, sequential, session-based, bundle, cross-domain, and multi-behavior recommendation.These scenarios differ in inputs and relationships, such as social influence, historical sequences, anonymous sessions, bundles, domains, or behavior types.
  • Scenarios: Session-based recommendation predicts the next item from an anonymous short session when long-term user profiles and histories are unavailable.Subsequent sessions from the same user are handled independently in this setting.
  • Scenarios: Bundle recommendation recommends item combinations, while multi-behavior recommendation models multiple interactions such as clicks, carts, sharing, collecting, and purchases.Bundles introduce affiliation relations between bundles and their constituent items; multi-behavior tasks often represent behavior types as heterogeneous graph edges.
  • Objectives: Accuracy is the primary objective, while diversity, explainability, and fairness represent important beyond-accuracy objectives.Individual-level diversity concerns topic coverage within a user’s list, whereas system-level diversity concerns differences across users and exposure of long-tail items.
  • Applications: Recommender systems support applications including product recommendation and point-of-interest recommendation, which account for business, spatial, and temporal factors.E-commerce models may consider adding-to-cart and purchasing, while POI recommendation concerns users’ next visits to locations.

2.2 Graph Neural Networks

Graph neural networks model graph-structured data by iteratively propagating and aggregating neighborhood embeddings. The section surveys graph construction, spectral and spatial designs, representative architectures, and optimization for downstream graph tasks.

  • Graph representation: GNNs process graph data whose nodes and edges represent entities and relations, but non-structured data requires manually defining this graph representation.Graph construction therefore depends either on pre-existing relational graphs or abstraction from data such as images and text.
  • Graph representation: Graphs may be homogeneous, heterogeneous, or hypergraphs, distinguished by the types of nodes and edges and by whether edges connect more than two nodes.These categories motivate specialized GNN designs for different graph structures.
  • GNN model categories: Spectral models filter graph signals in the Fourier-transformed spectral domain, whereas spatial models directly aggregate localized neighborhood features on graph structures.Despite different formulations, both approaches iteratively collect neighborhood information to capture high-order correlations.
  • Representative models: GCN approximates graph convolution with a first-order filter, while GraphSAGE samples and aggregates neighbors before merging their information with the target node.GraphSAGE aggregation can use options including MEAN and LSTM.
  • Representative models: GAT assigns different propagation weights through attention, and HetGNN is tailored to heterogeneous graphs by dividing neighbors into type-based subsets.The comparison of typical models encodes graph types, node and edge types, and neighborhood weights using visual attributes.
  • Network depth: Propagation is repeated L times to capture high-order structure, but most cases use L≤4 because deeper networks suffer from over-smoothing.Over-smoothing makes updated embeddings exhibit only small fluctuations as propagation depth increases.
  • Model optimization: GNN optimization maps learned node or edge embeddings to task targets and trains them with losses defined from graph labels or relationships.Examples include similarity-based link prediction, pair-wise positive-versus-negative discrimination, and point-wise classification losses such as cross-entropy.

2.3 Why are GNNs required for recommender systems

GNNs are applied to recommender systems because recommendation depends on graph-structured interactions and collaborative signals beyond direct neighbors. Their iterative propagation captures multi-hop connectivity and can use sparse semi-supervised signals during representation learning.

  • Motivation: GNN-based recommenders benefit from structural data, high-order connectivity, and enhanced supervision signals, contributing to state-of-the-art performance across recommendation settings.The survey identifies these as three perspectives explaining the success of GNN-based recommenders.
  • High-order connectivity: GNNs represent collaborative filtering as multi-hop graph neighborhoods, allowing user representations to incorporate items interacted with by similar users beyond first-order interactions.Traditional approaches mainly capture directly connected interaction records, while embedding propagation and aggregation incorporate higher-order connectivity.
  • Supervision signal: GNN-based models can leverage semi-supervised signals during representation learning to address sparse supervision in collected recommender-system data.The supplied passage identifies sparse supervision and semi-supervised learning as part of the motivation for GNN recommenders.

3 CHALLENGES OF APPLYING GNNS TO RECOMMENDER SYSTEMS

Applying GNNs to recommender systems requires coordinated decisions about graph construction, propagation and aggregation, optimization, and computational efficiency. These choices determine how recommendation tasks are represented, learned, and deployed at scale.

  • GNN-based recommendation requires four design decisions: graph construction, information propagation and aggregation, model optimization, and computation efficiency.These challenges are explicitly identified as central to applying GNNs in recommender systems.
  • Graph construction: Standard collaborative filtering can be represented as user-item link prediction on a bipartite graph, with users and items as nodes and observed interactions as edges.Graph construction includes both representing the input data and reformulating the recommendation objective as a graph task.
  • Graph construction: Node definitions affect model scale and parameter allocation, while edge definitions influence propagation, aggregation, graph quality, and optimization.Node types and feature representations also require task-specific decisions.
  • Propagation and aggregation: Propagation paths model high-order similarity, but aggregation functions must be chosen for each task because no single function performs best across tasks or datasets.Choices such as mean pooling, LSTM, max, and min also have different computational costs.
  • Optimization: Recommendation losses can be reformulated as graph-learning losses, while sampling strategies depend on graph structure and multiple tasks require balancing their objectives.Examples include point-wise link-prediction loss and BPR loss.
  • Computation efficiency: GNN recommendation is substantially more computationally expensive than NCF or FM, so neighbor sampling and graph pruning can improve scalability at the cost of recommendation performance.The cost increases with complex matrix operations and stacked GCN layers; spatial models such as PinSage are easier to deploy at scale.

4.1 Taxonomy

The survey organizes GNN-based recommendation research by recommendation stage, scenario, objective, and application. Stages describe the platform workflow, while scenarios capture different recommendation settings.

  • The taxonomy categorizes GNN-based recommendation research by stage, scenario, objective, and application.
  • Stages: Recommendation stages comprise matching for candidate selection, ranking for user-preference modeling, and re-ranking for criteria beyond accuracy.
  • Scenarios: Recommendation scenarios include social, sequential, and cross-domain recommendation.The supplied passage introduces these scenarios as part of the survey taxonomy.

4.2 GNN in Different Recommendation Stages

GNNs are applied across matching, ranking, and re-ranking, with each stage imposing different structural and efficiency requirements. Matching emphasizes scalable candidate generation, ranking emphasizes feature interactions, and re-ranking must combine multiple goals.

  • Matching: Matching selects hundreds of candidates from item pools containing millions, making efficiency essential and favoring coarse-grained preference modeling.Strict latency limits also constrain model complexity and feature usage.
  • Matching: Matching-stage GNNs usually perform embedding matching on user-item bipartite graphs and can balance candidate-generation accuracy with efficiency.Sampling supports large graphs, while recent methods simplify architectures or compare favorably with conventional collaborative-filtering methods.
  • Ranking: Ranking uses more accurate models and additional features, but its central GNN challenge is constructing structures that capture feature interactions.Fi-GNN uses a weighted fully connected feature graph, while PUP uses a heterogeneous graph for price-aware preference modeling.
  • Re-ranking: Re-ranking must account for item relationships such as substitutability and complementarity while personalizing recommendations to different users.GNNs provide a unified representation for item relationships and user preferences.
  • Re-ranking: Existing GNN-based re-ranking work considers only a few re-ranking goals, leaving other objectives less explored.

4.3 GNN in Different Recommendation Scenarios

GNN-based recommenders span social, sequential, session-based, bundle, and cross-domain scenarios. Across these settings, methods adapt graph construction and propagation to capture relations, behavior patterns, and information across entities or domains.

  • Social recommendation: Social recommendation combines social influence with user-item interaction preferences, either through separate graphs or a unified graph.The field has progressed from modeling only social networks toward jointly modeling social relations and user interactions.
  • Sequential recommendation: Sequential recommendation uses graphs to capture higher-order relations across user sequences, but current methods still rely heavily on RNNs or Transformers for sequence modeling.Graph construction may incorporate neighboring items, attributes, location, timestamps, and occurrence frequency.
  • Session-based recommendation: Session-based recommendation models short sessions with directed graphs to capture item transitions, while incorporating other sessions, hypergraphs, knowledge graphs, or additional links to enrich sparse context.The average Tmall session length is reported as 6.69, limiting the nodes and edges available from an individual session.
  • Bundle recommendation: Bundle recommendation must model item-bundle affiliation, sparse user-bundle interactions, and high-order relations.Early approaches jointly learn from user-item and user-bundle interactions, using parameter sharing or joint learning.

4.4 GNN for Different Recommendation Objectives

GNN-based recommendation addresses objectives beyond accuracy, including diversity, explainability, and fairness, through graph structure, propagation, and attention mechanisms. However, diversity and fairness remain constrained by weak signals, causal ambiguity, and accuracy–fairness trade-offs.

  • Diversified recommendation: GNN-based diversity methods target both individual-level topic coverage and system-level exposure of relevant long-tail items.Individual diversity requires covering varied user interests, while system diversity emphasizes long-tail items with fewer training samples.
  • Diversified recommendation: V2HT constructs an item graph with external-knowledge correlations and propagates frequent-item information to strengthen long-tail recommendations.Its graph introduces four edge types connecting frequent and long-tail items, followed by stacked GCN layers.
  • Diversified recommendation: FH-GAT balances accuracy and diversity with heterogeneous interaction graphs and neighbor-similarity loss, while NN/FN GCNs separate accuracy from weak-signal enhancement.The NN graph guarantees accuracy, whereas the FN graph enhances signals from diverse items.
  • Explainable recommendation: TriRank improves explainability by modeling users, items, and review-derived aspects as ternary relations in a heterogeneous graph.The model casts recommendation as a ternary relation ranking task and explicitly represents aspects in reviews.
  • Explainable recommendation: Knowledge-graph paths, meta-paths, and attention mechanisms provide explanations, but meta-path discovery can require domain knowledge and existing explanations may ignore causality.The survey notes that ignoring causality can produce wrong explanations.
  • Fair recommendation: Graph structures can magnify user unfairness, and existing fairness-oriented GNN methods generally incur some performance drop.Fairwalk, adversarial approaches, and FairGo reduce sensitive information or alter neighborhood sampling, but the accuracy–fairness trade-off remains unresolved.

4.5 GNN for Specific Recommendation Applications

GNNs are applied to specialized recommendation settings, including large-scale e-commerce, while other applications are organized under the survey’s stage, scenario, and objective taxonomy.

  • E-commerce recommendation: E-commerce recommendation studies use stacked GNN modules, deterministic clustering, and item-relation topology to address application-specific needs.These approaches target efficiency and representation of product relationships in large-scale commercial settings.
  • Scope of applications: The survey omits additional application-specific papers when they are already covered by corresponding stages, scenarios, or objectives.This keeps the application discussion focused on cases not sufficiently represented by the broader taxonomy.

5 OPEN PROBLEMS AND FUTURE DIRECTIONS

Future work must make GNN-based recommendation deeper, more scalable, more dynamic, and better adapted to diverse scenarios. The survey also highlights knowledge-graph integration, conversational interaction, and automated architecture design as promising directions.

  • Deep GNNs: Very deep GNNs should capture higher-order connectivity without over-smoothing, while delivering a performance gain over shallow models at acceptable computation and time costs.The survey notes that no universal strategy yet constructs very deep GNNs like CNNs.
  • Dynamic recommendation: Dynamic GNNs are needed because recommendation data, user preferences, and platform entities evolve over time.Sequential and session-based scenarios naturally collect data dynamically, while new users and products can enter the platform.
  • Computation efficiency: Neighbor sampling and subgraph updates improve scalability, but large-scale ranking remains difficult because thousands of features create complicated heterogeneous graphs.GraphSAGE-based sampling has enabled efficient embedding computation at billion-item scale, yet ranking-stage complexity remains a challenge.
  • Enhanced supervision: Sparse interaction supervision motivates self-supervised auxiliary tasks and contrastive training with graph-based data augmentations.Node dropout is one augmentation used to generate sample pairs for contrastive learning.
  • Generalization and AutoML: Because recommendation scenarios differ substantially, no single GNN architecture generalizes across all settings; AutoML can search large design spaces to reduce manual effort.The search space includes neighbor samplers, aggregators, interaction functions, and combinations of GNN layers.
  • Knowledge-graph enhancement: Knowledge graphs enrich user–item graphs and support cold-start, dynamicity, sequential recommendation, and explainability, but user-centric knowledge and complex user–item relations remain under-modeled.Future work is encouraged to extend knowledge-graph use toward diversity and fairness, which require more external knowledge.
  • Conversational recommendation: Conversational recommendation offers a way to collect preference information dynamically through direct user interaction.Users can explicitly convey consumption preferences while chatting with the system.

6 CONCLUSION

The survey synthesizes the rapid development of GNN models for recommender systems and presents their challenges, methods, and future directions. It is intended to support both junior and experienced researchers.

  • Conclusion: The survey systematically presents the challenges, methods, and future directions of GNN-based recommender systems.It covers both the field’s development history and recent advances.
Loading 2109.12843v3…