Source-linked AI summary
Graph Learning based Recommender Systems: A Review
Shoujin Wang, Liang Hu, Yan Wang, Xiangnan He, Quan Z. Sheng, Mehmet A. Orgun, Longbing Cao, Francesco Ricci, Philip S. Yu
TL;DR
Recommender systems must model complex, interconnected user, item, and attribute data, while interaction sparsity and heterogeneous relations create important challenges. This paper systematically formalizes and reviews graph learning based recommender systems, organizing their data characteristics, challenges, techniques, datasets, and research directions. It concludes that GLRS provide a comprehensive framework for extracting relational knowledge relevant to recommendation, with efficiency and relation modeling remaining important boundaries.
Problem
Recommender-system objects are interconnected through complex relations, while interaction data is often sparse and insufficient to capture preferences and item characteristics.
Method
The paper provides a systematic review that formalizes GLRS, categorizes graph-related challenges, surveys graph learning techniques and datasets, and discusses open research directions.
Results
The review organizes GLRS progress around graph representations and learning approaches that extract relational knowledge for recommendation, including methods addressing social influence, side information, and sequential interactions.
Takeaways & Limitations
GLRS offers a framework for modeling interconnected recommendation data and can support recommendation accuracy, explainability, and the use of heterogeneous relations.
Takeaways & Limitations
Random-walk-based recommender systems can be inefficient because they score all candidate items at each step and lack optimizable model parameters.
Abstract
from arXiv · showhide
Recent years have witnessed the fast development of the emerging topic of Graph Learning based Recommender Systems (GLRS). GLRS employ advanced graph learning approaches to model users' preferences and intentions as well as items' characteristics for recommendations. Differently from other RS approaches, including content-based filtering and collaborative filtering, GLRS are built on graphs where the important objects, e.g., users, items, and attributes, are either explicitly or implicitly connected. With the rapid development of graph learning techniques, exploring and exploiting homogeneous or heterogeneous relations in graphs are a promising direction for building more effective RS. In this paper, we provide a systematic review of GLRS, by discussing how they extract important knowledge from graph-based representations to improve the accuracy, reliability and explainability of the recommendations. First, we characterize and formalize GLRS, and then summarize and categorize the key challenges and main progress in this novel research area. Finally, we share some new research directions in this vibrant area.
1 Introduction
GLRS applies graph learning to recommendation because users, items, attributes, and other objects are connected by complex relations. The paper formalizes and systematically reviews this emerging area, including its challenges, progress, and open directions.
- Motivation: Graph learning is well suited to recommender systems because it can learn complex relations among users, items, attributes, and context.Random walks and graph neural networks are identified as techniques for learning relations represented in graphs.
- Formalization: Graph Learning based Recommender Systems (GLRS) model recommendation data as graphs whose nodes represent objects and whose edges represent relations.The model learns from graph topology and content information to produce recommendation results.
- Formalization: Depending on the recommendation scenario, GLRS may use homogeneous sequences or heterogeneous networks and predict ratings or item rankings.The objective may maximize utility or the probability of forming links between nodes.
- Contributions: The review systematically analyzes GLRS graph challenges, categorizes technical progress, and discusses open research directions.Its stated contributions cover data-driven challenge analysis, state-of-the-art organization, and future research references.
2 Data Characteristics and Challenges
RS data includes interactions and side information that can be represented through several graph structures. These representations address sparsity, preference modeling, and relational complexity, while creating challenges for effective information propagation and aggregation.
- Data Types: RS data comprises user-item interactions and side information, with interactions further divided into sequential and general forms.Interaction data may be explicit, such as ratings, or implicit, such as clicks and views.
- General Interaction Data: General interaction matrices map naturally to user-item bipartite graphs, with weighted edges for ratings and unweighted edges for implicit interactions.Users and items form the two node parts, while interactions connect nodes across the parts.
- General Interaction Data: Bipartite graph propagation can enrich sparsely connected users and items, mitigating sparsity and cold-start issues faced by matrix completion.The remaining challenge is propagating information between users and items effectively and efficiently.
- Sequential Interaction Data: Sequential interaction datasets record time-ordered user-item interactions and can be represented as directed graphs whose paths encode interaction order.Multi-type sequences represent each element as an interaction-type and item pair, producing compound nodes.
- Sequential Interaction Data: Directed graph representations capture complicated interaction transitions that deviate from simple consecutive time-series patterns.Neighbour aggregation and multidirectional connections help model repeated or complex transitions in sequences.
- Side Information: Heterogeneous graphs combine interaction and attribute information, enabling cross-type propagation but requiring selective aggregation of useful attributes.The selection problem remains important for improving recommendation performance.
- Side Information: Social information forms a user graph whose combination with interaction data creates a heterogeneous graph for social recommendation.A central challenge is modeling how many neighbour orders to use and how strongly different neighbours influence each user.
- Side Information: Item ontology knowledge represents hierarchical item relations and supports understanding users’ multi-level preferences and recommendation explainability.Propagating user preferences through the hierarchy remains challenging.
3 Graph Learning Approaches for RS
The paper categorizes graph-learning approaches for recommender systems into technical families, including random-walk, graph-embedding, and GNN-based methods. These approaches learn graph relations to support recommendation, while differing in representations, architectures, and efficiency trade-offs.
- The technical categorization divides GLRS into random-walk, graph-embedding, and GNN-based approaches.Graph embedding methods are further divided into GFRS, GDRRS, and GNERS, while GNN-based methods include GATRS, GGNNRS, and GCNRS.
- Random Walk Approach: Random-walk recommenders propagate preference or interaction information through graph transitions and rank nodes reached after several steps.Random walks can capture complex, higher-order, and indirect relations, while restart variants preserve the context of the starting node.
- Random Walk Approach: Random-walk methods require ranking scores for all candidate items at each step and lack parameters optimized for the recommendation objective.These properties make them less efficient and distinguish them from most learning-based paradigms.
- Graph Embedding Approach: Graph embedding maps graph nodes to low-dimensional vectors that encode structural information for downstream recommendation tasks.GFRS factorize inter-node commuting matrices based on meta-paths, GDRRS learn distributed representations from adjacent relations, and GNERS use neural networks for embeddings.
- GNN Approach: GNN-based recommenders use neural networks to learn informative graph representations and are categorized as GATRS, GGNNRS, or GCNRS.The taxonomy includes attention-based, gated, and convolutional architectures, with GraphSage included under GCN-based methods.
4 GLRS Algorithms and Datasets
The survey collects open-source GLRS algorithms and publicly accessible real-world datasets to support empirical analysis. The resources span varied inputs, learning approaches, tasks, domains, and dataset characteristics.
- Table 2 summarizes representative open-source GLRS algorithms selected because they are commonly used as baselines.The algorithms accept varied input data and use different learning approaches for different learning tasks.
- Table 3 lists public, real-world datasets from multiple domains with different characteristics for evaluating GLRS algorithms.These datasets are presented as resources for empirical analysis of the surveyed approaches.
5 Open Research Directions
The survey identifies open directions by matching persistent GLRS challenges with existing research progress. Proposed directions address temporal evolution, causal explainability, cross-domain structure, and scalable online processing.
- Self-evolutionary RS with dynamic-graph learning: Dynamic-graph learning could support self-evolutionary recommender systems because users, items, and their interactions change over time.The paper notes that graph dynamics can directly affect user and requirement modeling and recommendation results, but remain underestimated.
- Explainable RS with causal graph learning: Causal graph learning is proposed for explainable recommendation because understanding the reasons and intents behind user choices remains incomplete.The paper presents this understanding as a critical step toward reliable and explainable recommendations.
- Cross-domain RS with multiplex graph learning: Multiplex graph learning is proposed for cross-domain recommendation when interactions come from multiple intercorrelated sources, systems, or modalities.Multiplex networks represent these interactions across layers whose nodes may or may not connect between layers.
- High-efficiency online RS with large-scale graph learning: Large-scale GLRS requires more efficient algorithms because graph-structured data increases the time and space costs of processing and recommendation learning.The paper calls for faster online processing and continual model updates to generate timely recommendations.
6 Conclusions
Graph Learning based Recommender Systems (GLRS) are presented as a promising next-generation recommender-system paradigm. The review offers an overview of their recent progress, challenges, and future research directions for academia and industry.
- Graph Learning based Recommender Systems (GLRS) are presented as a promising next-generation recommender-system paradigm.
- The review provides an overview of recent GLRS progress, challenges, and future research directions.
- The review is intended to support both academic and industry communities.