Global Context Enhanced Graph Neural Networks for Session-based Recommendation

Ziyang Wang, Wei Wei, Gao Cong, Xiao-Li Li, Xianling Mao, Minghui QiuOpenReview 2020 · DOI · OpenReview

TL;DR

This paper introduces GCE-GNN, a novel graph neural network approach for session-based recommendation. It tackles the limitation of existing methods by leveraging item transition patterns from all past sessions, not just the current one, to build richer item representations and improve next-item predictions.

What Problem Does This Paper Address?

Session-based recommendation (SBR) aims to predict the next item a user might be interested in, given a sequence of their recent interactions within a single session. This is crucial for scenarios where user identities are unknown or short-lived, like anonymous browsing on e-commerce sites or streaming platforms. The core challenge is that most SBR models focus solely on the current session's item transitions. This overlooks valuable contextual information present in other past sessions, which could contain relevant patterns or help disambiguate noisy interactions within the current session. Simply lumping all past sessions together can also introduce irrelevant information, hindering performance.

Methods

GCE-GNN proposes a two-pronged approach to item representation learning:

  • Session Graph: This component uses a standard Graph Neural Network (GNN) on a graph constructed from the current session's item transitions. This captures the immediate sequential patterns within the session.
  • Global Graph: This is the paper's main innovation. It constructs a "global graph" by considering pairwise item transitions across all historical sessions. To avoid noise, it uses an "ε-neighbor set" to define relationships and a session-aware attention mechanism to selectively incorporate relevant neighbor information from the global graph. This global context is then used to learn global-level item embeddings.

Finally, GCE-GNN combines the session-level and global-level item representations using a soft attention mechanism to generate a final session representation, which is then used to predict the next item. They also introduce a reversed position embedding to better capture the importance of items within a session sequence.

Datasets

The authors evaluated their method on three real-world datasets:

  • Diginetica: From the CIKM Cup 2016, representing typical transaction data.
  • Tmall: From the IJCAI-15 competition, containing anonymized shopping logs.
  • Nowplaying: Describing music listening behavior.

These datasets were preprocessed by filtering short sessions and infrequent items, and then split into training and testing sets.

Key Insights and Ideas

  • Global Context is Key: The most significant insight is that item transition patterns from all past sessions, not just the current one, are crucial for robust SBR. This global context helps disambiguate noisy current sessions and provides a richer understanding of user preferences.
  • Subtle Global Integration: The paper emphasizes the importance of how global context is integrated. GCE-GNN's session-aware attention mechanism on the global graph is designed to selectively incorporate relevant information, preventing the dilution of signal by irrelevant past interactions.
  • Reversed Position Matters: The authors found that using reversed position embeddings (counting from the end of the session) is more effective than forward position embeddings for SBR. This is because the items closer to the end of a session are often more indicative of the user's immediate intent.

Key Results

GCE-GNN consistently outperformed nine baseline methods, including state-of-the-art approaches like SR-GNN and FGNN, across all three datasets and evaluation metrics (P@10, P@20, MRR@10, MRR@20). For instance, on Diginetica, GCE-GNN achieved a P@20 of 54.22%, a significant improvement over SR-GNN's 51.26%. The ablation studies confirmed the effectiveness of both the global graph and the session-aware attention mechanism.

Limitations and Future Work

  • Hyperparameter Sensitivity: The paper notes that the performance can be sensitive to hyperparameter settings, particularly the dropout ratio, and that finding the optimal balance is important to avoid overfitting.
  • Global Graph Construction: While the ε-neighbor set helps, the paper acknowledges that the global graph construction might still introduce some noise. Future work could explore more sophisticated methods for filtering irrelevant global transitions.
  • Computational Cost: Building and processing the global graph can be computationally intensive, especially with very large datasets. Future research could focus on more efficient graph construction and propagation techniques.

Summary updated

Abstract

Session-based recommendation (SBR) is a challenging task, which aims at recommending items based on anonymous behavior sequences. Almost all the existing solutions for SBR model user preference only based on the current session without exploiting the other sessions, which may contain both relevant and irrelevant item-transitions to the current session. This paper proposes a novel approach, called Global Context Enhanced Graph Neural Networks (GCE-GNN) to exploit item transitions over all sessions in a more subtle manner for better inferring the user preference of the current session. Specifically, GCE-GNN learns two levels of item embeddings from session graph and global graph, respectively: (i) Session graph, which is to learn the session-level item embedding by modeling pairwise item-transitions within the current session; and (ii) Global graph, which is to learn the global-level item embedding by modeling pairwise item-transitions over all sessions. In GCE-GNN, we propose a novel global-level item representation learning layer, which employs a session-aware attention mechanism to recursively incorporate the neighbors' embeddings of each node on the global graph. We also design a session-level item representation learning layer, which employs a GNN on the session graph to learn session-level item embeddings within the current session. Moreover, GCE-GNN aggregates the learnt item representations in the two levels with a soft attention mechanism. Experiments on three benchmark datasets demonstrate that GCE-GNN outperforms the state-of-the-art methods consistently.
BibTeX
@misc{wang2020global,
  author = {Ziyang Wang and Wei Wei and Gao Cong and Xiao-Li Li and Xianling Mao and Minghui Qiu},
  title = {{Global Context Enhanced Graph Neural Networks for Session-based Recommendation}},
  year = {2020},
  eprint = {2106.05081},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.05081},
  howpublished = {OpenReview},
}
Figure