Source-linked AI summary
TAGNN: Target Attentive Graph Neural Networks for Session-based Recommendation
Feng Yu, Yanqiao Zhu, Qiang Liu, Shu Wu, Liang Wang, Tieniu Tan
TL;DR
Session-based recommendation must predict actions from anonymous sessions despite complex item transitions and diverse interests that fixed session representations do not capture. TAGNN combines graph neural networks with target-aware attention to model transitions and activate interests relevant to each candidate item. Experiments on real-world benchmark datasets demonstrate its effectiveness and state-of-the-art performance.
Problem
Session-based recommenders need to predict actions without long-term user profiles, while fixed session vectors overlook complex transitions and diverse, target-dependent interests.
Method
TAGNN combines session-graph modeling and graph neural networks with target-aware attention over historical session items for each candidate target.
Results
Experiments on real-world datasets show that TAGNN achieves state-of-the-art performance and demonstrates effectiveness over benchmark methods.
Takeaways & Limitations
Target-aware representations jointly account for candidate-specific user interests and complex item transitions in session-based recommendation.
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
Session-based recommendation predicts actions from anonymous in-session behavior, but existing approaches overlook complex transitions and target-specific interests. TAGNN addresses these issues with session graphs, graph neural networks, and target-aware attention, achieving state-of-the-art performance in experiments.
- Session-based recommendation predicts a user’s next action from previous behaviors when long-term profiles or identities may be unavailable.
- Sequence-based methods model only consecutive-item transitions, making complex patterns such as repeated purchases difficult to capture.The session v1 → v2 → v1 → v3 illustrates a to-and-fro relationship that sequence methods may miss.
- Representing every session with one fixed vector cannot express diverse user interests and limits model expressiveness.Increasing the vector dimension may risk overfitting and deteriorate performance.
- Target-aware attention activates historical interests according to their relevance to each target item, producing target-specific session representations.The module focuses attention on behaviors related to the candidate item being considered.
- TAGNN constructs session graphs and applies graph neural networks to capture complex item transitions and obtain item embeddings.
- Extensive experiments on real-world datasets show that TAGNN achieves state-of-the-art performance.
2 THE PROPOSED METHOD: TAGNN
TAGNN represents sessions as directed graphs, learns item embeddings with GGNNs, and builds target-specific session representations for recommendation. Its target-aware attention activates interests relevant to each candidate item, while local and global preferences are combined with that target embedding.
- 2.1 Problem Statement and Constructing Session Graphs: TAGNN represents each anonymous session as a directed graph whose nodes are items and whose edges encode consecutive visits.The adjacency representation reflects outgoing and incoming connections between items.
- 2.2 Learning Item Embeddings: GGNNs propagate information across neighboring session-graph nodes to learn item vectors that capture complex item connections and transitions.The update and reset gates determine which information is preserved or discarded.
- 2.3 Constructing Target-Aware Embeddings: Target-aware attention computes a separate interest representation for each candidate target item by weighting historical session items according to their relevance.The resulting target embedding varies with different target items.
- 2.4 Generating Session Embeddings: TAGNN combines the last-visited item as a local preference, an attention-based global preference, and the target embedding into a target-specific session embedding.A linear transformation projects the concatenated vectors into one embedding space, producing different session embeddings for each target item.
- 2.5 Making Recommendation: Recommendation scores are computed from item–session inner products, converted into probabilities, and used to select the top-k candidate items.The model is trained with cross-entropy against one-hot ground-truth items using back-propagation through time.
3 EXPERIMENTS
The experiments evaluate TAGNN against representative baselines and test alternative session-embedding strategies on two real-world datasets. TAGNN achieves state-of-the-art performance, while ablations favor combining local embeddings with target-aware attentive global embeddings.
- Experimental Configurations: Experiments use Yoochoose2 and Diginetica, comparing TAGNN with frequency-, similarity-, factorization-, RNN-, attention-, and GNN-based baselines.Sessions with fewer than two items and items appearing fewer than five times are removed; evaluation uses Precision@20 and MRR@20.
- Comparison with Baseline Methods (RQ1): TAGNN achieves state-of-the-art performance on all datasets in both Precision@20 and MRR@20.The comparison reports these metrics in Table 1, with the highest performance highlighted.
- Comparison with Baseline Methods (RQ1): Deep learning methods greatly outperform conventional models, while Item-KNN surpasses most Markov-chain-based methods such as BPR-MF and FPMC.The text attributes the deep-learning advantage to stronger capacity for capturing complex user behavior and criticizes limited sequential dependencies.
- Ablation Studies (RQ2): The hybrid TAGNN-L+Att embedding strategy achieves the best performance on all datasets among the tested session representations.The ablation compares local-only, average-pooled global, attentive-global, and hybrid local-plus-attentive-global embeddings.
- Ablation Studies (RQ2): Combining attentive global and local embeddings captures both long- and short-term preferences, whereas average pooling performs poorly.The last item also has a strong impact on the user’s final action in the reported ablation.
4 CONCLUSIONS
The paper develops TAGNN for session-based recommendation by combining session-graph modeling with target-aware attention. Experiments on real-world benchmark datasets demonstrate the model’s effectiveness.
- Conclusions: TAGNN jointly models complex item transitions in sessions and user interests conditioned on a target item.The approach combines graph modeling with a target-aware attention module.
- Conclusions: Extensive empirical evaluation on real-world benchmark datasets demonstrates the effectiveness of TAGNN.