TAGNN: Target Attentive Graph Neural Networks for Session-based Recommendation

Feng Yu, Yanqiao Zhu, Qiang Liu, Shu Wu, Liang Wang, Tieniu TanSIGIR 2020 · DOI · OpenReview

Abstract

Session-based recommendation nowadays plays a vital role in many websites, which aims to predict users' actions based on anonymous sessions. There have emerged many studies that model a session as a sequence or a graph via investigating temporal transitions of items in a session. However, these methods compress a session into one fixed representation vector without considering the target items to be predicted. The fixed vector will restrict the representation ability of the recommender model, considering the diversity of target items and users' interests. In this paper, we propose a novel target attentive graph neural network (TAGNN) model for session-based recommendation. In TAGNN, target-aware attention adaptively activates different user interests with respect to varied target items. The learned interest representation vector varies with different target items, greatly improving the expressiveness of the model. Moreover, TAGNN harnesses the power of graph neural networks to capture rich item transitions in sessions. Comprehensive experiments conducted on real-world datasets demonstrate its superiority over state-of-the-art methods.

Introduction

Session-based recommendation (SBR) is the task of predicting a user's next action based solely on their current, often anonymous, session. This is crucial for platforms where users do not log in or lack long-term profiles. Existing SBR models face two major limitations: oversimplified transitions and the "fixed vector" bottleneck. Many models treat sessions as simple sequences, failing to capture complex, non-linear patterns like clicking back and forth between items. Furthermore, compressing an entire session into a single, fixed-length embedding vector fails to account for how a user's interest "activates" differently depending on the specific candidate item being considered.

Methodology

The TAGNN model addresses these limitations by transforming raw clicks into intelligent, target-specific predictions through several sophisticated stages:

  • Session Graph Construction: Sessions are modeled as directed graphs where nodes are items and edges represent consecutive clicks. This naturally captures structural relationships and repeated behaviors.
  • Learning Item Embeddings (GNN): The model employs Gated Graph Neural Networks (GGNNs) to allow information to flow between neighboring items in the session graph, creating context-aware embeddings.
  • Target-Aware Attention: This is the core innovation. For every potential "target" item, the model calculates an attention score against every item in the current session, creating a dynamic, target-specific interest representation.
  • Hybrid Session Embedding: The final session representation is a concatenation of a local embedding (the last item clicked), a global embedding (a general session summary), and the target-aware embedding.
The TAGNN architecture illustrating the flow from session graphs and GNNs to target-aware predictions.

Experimental Results

TAGNN was evaluated on two standard industry benchmarks: Yoochoose (e-commerce clicks) and Diginetica (transaction data). Both datasets were preprocessed to ensure data quality by removing very short sessions and infrequent items.

As shown in the table below, TAGNN consistently outperformed several strong baselines, including traditional methods (Item-KNN), RNN-based models (GRU4Rec, NARM), and the previous graph-based state-of-the-art (SR-GNN).

Performance comparison of TAGNN against state-of-the-art baselines on the Diginetica and Yoochoose 1/64 datasets.

Key insights from the results include:

  • Dynamic vs. Static: Making the session vector dependent on the target item significantly increases model expressiveness without increasing dimensionality.
  • Structural Understanding: GNNs effectively capture the messy, non-linear reality of user browsing journeys compared to rigid sequence models.
  • Relevance Filtering: The target-aware module identifies which past actions are actually relevant to the item currently being evaluated.

Conclusion

TAGNN demonstrates that adaptive interest activation is key to improving session-based recommendations. By moving beyond fixed session representations and leveraging the structural power of GNNs, the model achieves superior accuracy on real-world datasets. Future work may explore the computational latency of per-target attention in large-scale systems and the integration of long-term user profiles where available.

Summary updated

Abstract

Session-based recommendation nowadays plays a vital role in many websites, which aims to predict users' actions based on anonymous sessions. There have emerged many studies that model a session as a sequence or a graph via investigating temporal transitions of items in a session. However, these methods compress a session into one fixed representation vector without considering the target items to be predicted. The fixed vector will restrict the representation ability of the recommender model, considering the diversity of target items and users' interests. In this paper, we propose a novel target attentive graph neural network (TAGNN) model for session-based recommendation. In TAGNN, target-aware attention adaptively activates different user interests with respect to varied target items. The learned interest representation vector varies with different target items, greatly improving the expressiveness of the model. Moreover, TAGNN harnesses the power of graph neural networks to capture rich item transitions in sessions. Comprehensive experiments conducted on real-world datasets demonstrate its superiority over state-of-the-art methods.
BibTeX
@inproceedings{yu2020tagnn,
  author = {Feng Yu and Yanqiao Zhu and Qiang Liu and Shu Wu and Liang Wang and Tieniu Tan},
  title = {{TAGNN: Target Attentive Graph Neural Networks for Session-based Recommendation}},
  booktitle = {International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year = {2020},
  doi = {10.1145/3397271.3401319},
}
Figure