Source-linked AI summary
Rethinking the Item Order in Session-based Recommendation with Graph Neural Networks
Ruihong Qiu, Jingjing Li, Zi Huang, Hongzhi Yin
TL;DR
Session-based recommendation must infer shifting preferences from anonymous recent interactions, but time-ordered sequence models do not fully capture complicated item transitions. The paper proposes FGNN, which combines session-graph structure with sequence information through WGAT and Readout, achieving state-of-the-art results on benchmark e-commerce datasets.
Problem
Session-based recommendation must represent an anonymous user’s recent preference without long-term history, while item transitions are more complicated than straightforward chronological order.
Method
FGNN converts each session sequence into a weighted directed session graph, applies WGAT layers for item embeddings, and uses a Readout function to form a session representation for next-item recommendation.
Results
FGNN achieves state-of-the-art results on two benchmark e-commerce datasets, Yoochoose and Diginetica.
Takeaways & Limitations
Modeling latent item-transition order alongside sequence order provides the paper’s supported approach to representing session-level preference for recommendation.
Takeaways & Limitations
The paper identifies the simplified separation of long-term and short-term preference as insufficient for capturing complicated item-transition patterns.
Abstract
from arXiv · showhide
Predicting a user's preference in a short anonymous interaction session instead of long-term history is a challenging problem in the real-life session-based recommendation, e.g., e-commerce and media stream. Recent research of the session-based recommender system mainly focuses on sequential patterns by utilizing the attention mechanism, which is straightforward for the session's natural sequence sorted by time. However, the user's preference is much more complicated than a solely consecutive time pattern in the transition of item choices. In this paper, therefore, we study the item transition pattern by constructing a session graph and propose a novel model which collaboratively considers the sequence order and the latent order in the session graph for a session-based recommender system. We formulate the next item recommendation within the session as a graph classification problem. Specifically, we propose a weighted attention graph layer and a Readout function to learn embeddings of items and sessions for the next item recommendation. Extensive experiments have been conducted on two benchmark E-commerce datasets, Yoochoose and Diginetica, and the experimental results show that our model outperforms other state-of-the-art methods.
1 INTRODUCTION
Session-based recommendation addresses anonymous users by modeling recent interactions, but item preferences may follow an inherent transition order beyond chronology or simple attention.
- Recent interactions can reveal preference shifts that conventional content-based and collaborative-filtering recommenders miss when long-term history is unavailable.
- Sequence-based methods treat sessions as time series, although user preference does not completely depend on chronological order.
- Methods separating long-term and short-term preference may bias representations by assigning the last item to local preference and all earlier items to global preference.
- FGNN constructs a session graph, formulates next-item recommendation as graph classification, and learns item-transition order with WGAT and Readout.
- The model introduces WGAT for weighted neighbor information and Readout for an appropriate graph-level order, with state-of-the-art results on Yoochoose and Diginetica.
2 RELATED WORK
Related work spans general and session-based recommendation, sequential models, FGNN’s graph pipeline, and graph neural networks for representation learning and classification.
- General recommender systems commonly use collaborative filtering over complete user histories, but anonymous commercial settings make these methods fail.
- Session-based recommendation instead uses recent user-item interactions as a proxy for current preference without requiring all historical actions.
- Sequential recommendation models item dependence with Markov chains, probabilistic decision trees, or Markov decision processes to predict subsequent clicks.
- Deep learning approaches include recurrent models such as GRU4REC, attention-based STAMP, graph-based SR-GNN, and session-history modeling in SSRM.
- FGNN’s pipeline converts a session sequence into a graph, applies WGAT, aggregates node features with Readout, and compares the graph representation with the item set.
- GNN methods learn node or graph representations through information flow over graph edges, supporting graph-level tasks such as classification.
3 PRELIMINARIES
Graph neural networks learn representations from node features and graph structure through iterative neighborhood aggregation, while graph classification requires a permutation-invariant graph-level Readout.
- Graph classification learns a graph representation h_G and predicts a graph label from that representation.
- GNNs iteratively aggregate neighboring-node information so final node representations capture structural and feature information within k-hop neighborhoods.
- The aggregation function combines neighboring features, while the mapping function transforms self and neighborhood information into a new node vector.
- A Readout function aggregates final-layer node features into a graph-level representation for graph classification.
- The Readout function must be permutation invariant so graph representations do not depend on node ordering.
4 METHOD
FGNN converts each session sequence into a weighted directed session graph, applies weighted graph attention to learn item embeddings, and uses Readout to form a session representation for next-item recommendation.
- Session Graph: FGNN converts each session sequence into a weighted directed graph and formulates next-item recommendation as graph classification.The graph represents item transitions, with edge weights given by transition frequency and self-loops added where needed.
- Session Graph: The session graph contains item nodes initialized with embedding vectors and directed edges representing consecutive clicks.Each edge weight records how frequently that transition occurs within the session.
- Weighted Graph Attentional Layer: WGAT incorporates edge weights into attention aggregation so neighboring nodes can convey information through the weighted directed session graph.Attention coefficients determine how strongly neighboring nodes influence each node, and multiple WGAT layers capture higher-order transition patterns.
- Weighted Graph Attentional Layer: WGAT begins with item embedding features, applies shared linear mappings and neighbor attention, then combines multi-head outputs into item-level embeddings.The final node features are used as inputs to session embedding computation.
- Readout Function: Readout produces a graph-level session representation from node features while learning an order of reading suited to the session graph.The resulting representation is formed through query-based attention over node embeddings and recurrent updates.
5 EXPERIMENTS
Experiments evaluate FGNN on benchmark session-based recommendation datasets using ranking metrics and comparisons with traditional, neural, and graph-based baselines. FGNN achieves state-of-the-art results, while component studies support WGAT and the Readout function for modeling item-transition structure.
- Experimental Setup: Experiments use Yoochoose and Diginetica benchmark e-commerce datasets after filtering short sessions and infrequent items.Sessions of length 1 and items occurring fewer than five times were removed.
- Experimental Setup: R@20 and MRR@20 are the primary evaluation metrics, with top-5 and top-10 results also reported for detailed comparison.R@K measures whether the desired item appears among the top K recommendations, while MRR@K reflects its ranking position.
- General Comparison: FGNN outperforms all baseline methods on R@20 and MRR@20 across the three evaluated datasets, achieving state-of-the-art performance.The comparison includes traditional, neural-network, and graph-based session recommendation methods.
- General Comparison: FGNN also achieves the best top-5 and top-10 results on Yoochoose1/64.The authors attribute this performance to WGAT's node-level encoding and Readout's learned order of graph nodes.
- WGAT Analysis: WGAT outperforms other tested GNN layers because it captures edge direction and explicit weights while providing stronger representation learning than gated graph networks.The comparison reports results using R@20 and MRR@20.
- Readout Analysis: The Readout function learns an inherent item-transition order rather than relying only on time order or a hand-crafted long-term/short-term split.The paper links this learned order to more accurate session embeddings and recommendations.
6 CONCLUSION
The work addresses session-based recommendation without user history by using WGAT layers and a Readout function to represent session preferences, achieving state-of-the-art results on benchmark e-commerce datasets.
- WGAT layers learn item embeddings within a session, which the Readout function processes into a session embedding representing the user’s preference.
- The proposed method achieves state-of-the-art results on benchmark e-commerce datasets.
- Future work should use inter-session information to represent the user’s preference more accurately.