Source-linked AI summary
Session-based Recommendation with Graph Neural Networks
Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, Tieniu Tan
TL;DR
Session-based recommendation must predict next actions from anonymous sessions despite limited behavior for estimating users and overlooked complex item transitions. SR-GNN models sessions as graphs, captures item transitions with graph neural networks, and combines global preference with current interest in an attention-based session embedding. The paper reports consistent outperformance of state-of-the-art methods on real-world datasets.
Problem
Anonymous sessions provide limited behavior for accurate user representations, while prior sequential methods overlook complex transitions among session items.
Method
SR-GNN models session sequences as graphs, uses graph neural networks to capture item transitions, and forms session embeddings from global preference and current interest.
Results
SR-GNN consistently outperforms other state-of-the-art session-based recommendation methods in comprehensive real-world experiments.
Takeaways & Limitations
The method represents sessions from latent item vectors rather than relying on user representations and incorporates both complex transitions and session interests.
Takeaways & Limitations
Markov-chain baselines combine past components independently, an assumption that confines prediction accuracy.
Abstract
from arXiv · showhide
The problem of session-based recommendation aims to predict user actions based on anonymous sessions. Previous methods model a session as a sequence and estimate user representations besides item representations to make recommendations. Though achieved promising results, they are insufficient to obtain accurate user vectors in sessions and neglect complex transitions of items. To obtain accurate item embedding and take complex transitions of items into account, we propose a novel method, i.e. Session-based Recommendation with Graph Neural Networks, SR-GNN for brevity. In the proposed method, session sequences are modeled as graph-structured data. Based on the session graph, GNN can capture complex transitions of items, which are difficult to be revealed by previous conventional sequential methods. Each session is then represented as the composition of the global preference and the current interest of that session using an attention network. Extensive experiments conducted on two real datasets show that SR-GNN evidently outperforms the state-of-the-art session-based recommendation methods consistently.
1 Introduction
Session-based recommendation must make next-click predictions from limited anonymous behavior, where conventional methods struggle with user representations and complex item transitions. SR-GNN models sessions as graphs, uses graph neural networks and attention-based session embeddings, and outperforms state-of-the-art methods in real-world experiments.
- Motivation: Anonymous sessions provide limited behavior histories, making accurate user representations difficult for session-based recommendation.Conventional recommendation systems often assume recorded user profiles and past activities, which may be unavailable.
- Limitations of prior methods: Markov-chain methods rely on an independence assumption that confines prediction accuracy.
- Limitations of prior methods: RNN-based methods achieve promising results but have difficulty estimating users from limited anonymous session clicks.Their hidden vectors are commonly treated as user representations.
- Proposed method: SR-GNN models separated session sequences as graph-structured data to capture complex item transitions and generate accurate item latent vectors.
- Proposed method: SR-GNN represents each session by combining global preference and current interest through latent item vectors and an attention network.The workflow proceeds from session graphs to gated graph neural network node vectors, then to attention-based session representations and next-click probabilities.
- Results: Extensive experiments on real-world datasets show that SR-GNN evidently outperforms state-of-the-art session-based recommendation methods.
2 Related Work
Related work spans conventional recommendation, Markov-chain sequential models, RNN-based methods, and graph neural networks. These approaches differ in how they use item similarity, sequence order, user behavior, and graph structure.
- Conventional recommendation methods: Matrix factorization is poorly suited to session-based recommendation because it relies on positive clicks and conventional item neighborhoods may ignore sequence order.
- Markov-chain methods: Markov-chain methods predict next behavior from previous actions, but their independent combination of past components can constrain prediction accuracy.
- Deep-learning-based methods: RNN-based methods model session clicks and clicked-item features, with later work improving recurrent models through data augmentation and temporal-shift modeling.
- Neural networks on graphs: Graph neural networks generate representations for graph-structured data, while gated GNNs operate on directed graphs using gated recurrent units and back-propagation through time.
3 The Proposed Method
SR-GNN models anonymous session sequences as directed graphs, using gated graph neural networks to learn item representations and capture item transitions. It represents each session by combining current interest with global preference before scoring candidate items.
- Problem formulation: Session-based recommendation predicts the next clicked item from an anonymous session's current sequential data.
- Constructing session graphs: Each session becomes a directed graph whose nodes are items and whose edges encode consecutive clicks, with repeated transitions receiving normalized weights.
- Learning item embeddings: Gated graph neural networks propagate neighborhood information through the session graph and use update and reset gates to control retained information.
- Generating session embeddings: The method represents a session without a distinct user vector, combining the last-clicked item's local embedding with an attention-weighted global embedding.
- Generating session embeddings: A linear transformation compresses the concatenated local and global embeddings into a hybrid session representation used to score candidate items.
- Recommendation and training: The model converts candidate-item scores into next-click probabilities and trains each session graph with cross-entropy loss using BPTT.
4 Experiments and Analysis
Experiments compare SR-GNN with baselines, connection schemes, session representations, and session lengths across real-world datasets. SR-GNN consistently performs strongly, with graph modeling, hybrid embeddings, and attention contributing to robust recommendation behavior.
- Comparison with Baseline Methods: SR-GNN achieves the best P@20 and MRR@20 performance on all three evaluated datasets.The comparison uses P@20 and MRR@20 against state-of-the-art session-based recommendation methods.
- Comparison with Baseline Methods: Neural methods outperform conventional baselines, but SR-GNN further models complex item transitions that sequential models may ignore.NARM and STAMP outperform conventional methods, while SR-GNN models sessions as graphs to capture implicit connections between clicks.
- Comparison with Variants of Connection Schemes: All three connection schemes perform better than or nearly as well as STAMP and NARM, supporting graph-based session modeling.The variants are SR-GNN, normalized global connections, and full connections.
- Comparison with Variants of Connection Schemes: Global connections can reduce current-session edge influence and cause performance degradation, while explicitly adding all higher-order connections performs slightly worse than SR-GNN.The reported difference between SR-GNN and SR-GNN-FC is small, suggesting not every higher-order transition should become a direct connection.
- Comparison with Session Representations: The hybrid session embedding performs best, and attention outperforms average pooling by selecting significant behavior from potentially noisy sessions.The local-only variant also performs close to the attention-based method, indicating value in both current interest and long-term preference.
- Analysis on Session Sequence Lengths: SR-GNN and its variants remain stable across short and long sessions, whereas STAMP and NARM fluctuate or degrade as session length increases.On Yoochoose, NARM performs well for short sessions but drops quickly for longer sequences.
5 Conclusions
SR-GNN incorporates graph models into session representations to capture complex item transitions and combine long-term preferences with current interests. The paper reports consistent outperformance of other state-of-the-art methods, with performance evaluated across session lengths using P@20.
- SR-GNN incorporates graph models into session sequences to represent complex structures and item transitions.
- Table 3 evaluates different methods across session lengths using P@20.
- The method combines long-term preferences and current interests to predict users’ next actions.
- Comprehensive experiments report that SR-GNN consistently outperforms other state-of-the-art methods.