PaperlayerView exact revision on arXiv

Source-linked AI summary

GAG: Global Attributed Graph Neural Network for Streaming Session-based Recommendation

Ruihong Qiu, Hongzhi Yin, Zi Huang, Tong Chen

arXiv:2007.02747v2cs.IR

TL;DR

Streaming session-based recommendation must adapt to arriving sessions, while static training and some prior streaming approaches do not address the setting’s update and historical-interest needs. The paper combines a global-attributed session graph with a Wasserstein reservoir, and reports state-of-the-art performance on two real-world datasets.

  • Problem

    Most session-based recommendation research uses static batch training, while streaming sessions require prompt updates and memory-efficient retention of historical data.

    Source: §1 · p. 1

  • Method

    GAG represents a session as a graph with the user embedding as a global attribute, and a Wasserstein reservoir selects sessions by distance between recommendations and real interactions.

    Source: §1 · p. 2, §3.3 · p. 3, §3.6 · p. 5

  • Results

    Experiments on two real-world datasets report state-of-the-art performance for GAG and the Wasserstein reservoir.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The paper identifies cross-session information as a significant direction for future streaming session-based recommendation research.

    Source: §6 · p. 10

Abstract

from arXiv · show

Streaming session-based recommendation (SSR) is a challenging task that requires the recommender system to do the session-based recommendation (SR) in the streaming scenario. In the real-world applications of e-commerce and social media, a sequence of user-item interactions generated within a certain period are grouped as a session, and these sessions consecutively arrive in the form of streams. Most of the recent SR research has focused on the static setting where the training data is first acquired and then used to train a session-based recommender model. They need several epochs of training over the whole dataset, which is infeasible in the streaming setting. Besides, they can hardly well capture long-term user interests because of the neglect or the simple usage of the user information. Although some streaming recommendation strategies have been proposed recently, they are designed for streams of individual interactions rather than streams of sessions. In this paper, we propose a Global Attributed Graph (GAG) neural network model with a Wasserstein reservoir for the SSR problem. On one hand, when a new session arrives, a session graph with a global attribute is constructed based on the current session and its associate user. Thus, the GAG can take both the global attribute and the current session into consideration to learn more comprehensive representations of the session and the user, yielding a better performance in the recommendation. On the other hand, for the adaptation to the streaming session scenario, a Wasserstein reservoir is proposed to help preserve a representative sketch of the historical data. Extensive experiments on two real-world datasets have been conducted to verify the superiority of the GAG model compared with the state-of-the-art methods.

1 INTRODUCTION

Streaming session-based recommendation must update as sessions arrive, but static SR training is poorly suited to that setting. The paper proposes GAG with a Wasserstein reservoir to represent user interests and select sessions for streaming updates.

1 INTRODUCTION

  • Static SR models are unsuitable for streaming sessions because preferences change while memory and prompt retraining are constrained.Training on every new session is impractical, and typical SR models may not converge within the short update window.Source: §1 · p. 1
  • Conventional reservoirs store and sample discrete interactions, risking loss of session sequence information; online updates can overfit new data and forget historical preferences.Traditional reservoirs were designed to capture matrix-factorization information rather than session sequence patterns.Source: §1 · p. 2
  • Guo et al. use weighted session sampling, but require precomputed item features and combine SR with matrix factorization, limiting generalization and user-item correlation learning.Source: §1 · p. 2
  • GAG treats the user embedding as a session graph’s global attribute, while the Wasserstein reservoir selects informative training cases for streaming updates.It samples sessions according to the Wasserstein distance between recommendation lists and users’ real interactions; higher-distance sessions are treated as more informative for refinement.Source: §1 · p. 2
  • Experiments on two real-world datasets report state-of-the-art performance for GAG and the Wasserstein reservoir.Source: §1 · p. 2

2 RELATED WORK

Prior session-based recommendation work includes Markov models, recurrent and attention architectures, and graph-based approaches. Streaming recommendation research addresses updates and historical data retention, while graph-neural-network work introduces graph-level attributes.

2.1 Session-based Recommendation

  • Markov-chain approaches model item dependencies; later SR methods include GRU4REC, attention-based NARM and STAMP, GNN session encoders, and SSRM’s use of a user’s historical sessions.Source: §2.1 · p. 2

2.2 Streaming Recommendation

  • Streaming methods update models with new interactions, but may forget historical interactions; reservoir sampling stores long-term interactions, including in work applying it to streaming sessions.Source: §2.2 · pp. 2–3

2.3 Graph Neural Networks

  • Some GNN methods aggregate neighborhood information through message passing, while earlier work introduced global attributes to maintain graph-level features.Source: §2.3 · p. 3

3 METHOD

The method defines streaming session-based recommendation and introduces GAG, which combines a session graph with user information as a global attribute. Its graph-convolution and recommendation components are paired with a Wasserstein reservoir that prioritizes informative sessions for streaming updates.

3.1 Task Definition

  • At time t, SSR recommends the next item from a user’s session history while sessions arrive rapidly and computation resources are limited.Source: §3.1 · p. 3

3.2 Overview

  • GAG incorporates a user’s information as a global attribute in the session graph, alongside a Wasserstein reservoir for streaming-data learning.Source: §3.2 · p. 3

3.3 Global Attributed Session Graph

  • Each session becomes a weighted directed graph of its item nodes, edge weights, and associated user embedding as a global attribute.Source: §3.3 · p. 3

3.4 Global Attributed Graph Neural Network

  • GAG updates edge and node features before updating the global attribute; node updates aggregate incoming and outgoing neighborhoods and incorporate edge weights and the global attribute.Source: §3.4 · pp. 3–4

3.5 Recommendation

  • The updated global attribute represents the session graph, and item scores derived from the session and user representations are converted into recommendation probabilities.Source: §3.4 · p. 4, §3.5 · p. 4

3.6 Wasserstein Reservoir for Streaming Model

  • The Wasserstein reservoir samples sessions with higher probability when their predicted recommendation distributions are farther from real interactions, while new-user or new-item sessions are included directly.Source: §3.6 · p. 5

3.7 Training

  • Training treats recommendation as classification over the item set and uses cross-entropy between predicted and real interaction distributions.Source: §3.7 · pp. 5–6

4 EXPERIMENT SETUP

The experiments use two real-world datasets and a chronological streaming simulation, evaluating top-20 recommendations against representative baselines with Recall and MRR. The GAG model is trained with a shared embedding size and specified reservoir and window settings.

4.1 Dataset

  • The study uses 298,919 LastFM sessions and 198,680 Gowalla sessions after filtering sessions with fewer than 2 or more than 20 transactions.LastFM sessions group transactions from the same user over 8 hours; Gowalla sessions group check-ins within 1 day.Source: §4.1 · p. 6

4.2 Metrics

  • Data are split chronologically 60/40 into offline training and online candidate sets, with the candidate stream divided into five sequential test sets and model updates between tests.Each test set is presented in time order, and its updated model is used for the next test set.Source: §4.2 · p. 6

4.3 Baselines

  • Evaluation uses top-20 recommendations and compares models with Recall@K and MRR@K.The representative baselines include popularity methods, BPR-MF, GRU4REC, NARM, FGNN, and SSRM.Source: §4.2 · p. 6, §4.3 · p. 6

4.4 Training Detail

  • GAG uses 200-dimensional embeddings, one-layer MLPs, Adam with learning rate 0.003, batch size 100, and a reservoir of |D|/100.The window size is |C_new|/2 on each C ∪ C_new.Source: §4.4 · p. 6

5 EXPERIMENT RESULTS

The experiments assess GAG on two real-world datasets through four questions: its performance against existing methods, the global attribute's contribution, the Wasserstein reservoir, and parameter sensitivity.

Experiment scope

  • The experiments are designed to assess GAG's efficacy on two real-world datasets.The study frames the evaluation around four research questions.Source: §5 · p. 6

Research questions

  • The first research question compares GAG's performance with current state-of-the-art methods.This comparison is addressed in Section 5.1.Source: §5 · p. 6
  • The remaining questions examine the global attribute, Wasserstein reservoir performance, and GAG parameter sensitivity.These topics are addressed in Sections 5.2–5.4.Source: §5 · p. 6

5.1 Comparisons with Baseline Methods

GAG-50 outperforms all baselines across the reported overall situations, and the 200-dimensional GAG achieves state-of-the-art results. In the Gowalla top-5 and top-10 comparisons, GAG and GAG-50 retain their advantage.

5.1.1 General Comparison.

  • POP and S-POP perform much worse than the proposed model; POP misses session-specific behavior, while S-POP captures item re-occurrence patterns within a session.BPR-MF performs better than POP but does not outperform S-POP in SSR.Source: §5.1.1 · pp. 6–7
  • GAG-50 outperforms all baselines in every reported overall situation, while 200-dimensional GAG achieves state-of-the-art results.The overall comparison uses Recall@20 and MRR@20 on Gowalla and LastFM.Source: §5.1.1 · p. 6, §5.1 · p. 6, Figure 4 · p. 7
  • Attention-based NARM and SSRM improve over GRU4REC, and SSRM is the strongest baseline in the experiment.The authors attribute the improvement to attention's ability to learn item inter-dependencies.Source: §5.1.1 · p. 7
  • FGNN performs comparably to SSRM but worse than GAG-50, which the authors relate to FGNN's inability to model user information.Source: §5.1.1 · p. 7

5.1.2 In-depth Comparison.

  • On Gowalla, GAG and GAG-50 remain superior in top-5 and top-10 results; SSRM drops more than GNN-based methods at these cutoffs.The in-depth comparison uses Recall@K and MRR@K for K = 5 and 10.Source: Figure 5 · p. 7, §5.1.2 · p. 7
  • The discussion also reports that attention fails to distinguish item-transition patterns in sessions.Source: §5.1.2 · p. 7

5.2 Effect of Global Attribute

The experiments assess how the global attribute contributes to recommendation performance, using variants that place or remove it in different GAG update functions. Each use of the global attribute improves performance over FGNN, with node-update integration having the greater impact.

5.2 Effect of Global Attribute

  • The ablation evaluates Recall@20 and MRR@20 on Gowalla and LastFM.Source: §5.2 · p. 7

5.2.1 Ablation Study.

  • FGNN omits user information from both node updates and readout, while GAG-FGNN and GAG-NoGA retain it in different update functions.GAG-FGNN retains the global-attribute update; GAG-NoGA retains the attribute in the node update.Source: §5.2.1 · p. 7
  • Global-attribute use in the node update has a greater performance impact than its use in the global-attribute self-update.The comparison is between GAG-NoGA, which retains the attribute in node updates, and GAG-FGNN, which retains it in the global-attribute update.Source: §5.2.1 · pp. 7–8, Figure 6 · p. 8

5.3 Effect of Wasserstein Reservoir

The experiments test the Wasserstein reservoir’s sampling design and efficiency choices for streaming updates. Its full strategy performs best among the compared sampling variants, while reservoir and window size affect performance and storage or update workload.

5.3 Effect of Wasserstein Reservoir

  • The reservoir combines Wasserstein-distance-based sampling with direct inclusion of sessions containing new items or new users.The ablation compares the combined design with variants that isolate or omit these strategies.Source: §5.3 · p. 8
  • The Wasserstein reservoir achieves the best performance in all reported situations among the compared sampling strategies.The comparison includes static training, random sampling, direct addition of sessions with new items or users, and Wasserstein-based sampling.Source: §5.3 · p. 8

5.3.2 Reservoir Efficiency.

  • GAG performs best at reservoir size |D|/100, while SSRM reaches its highest performance at |D|/20, five times larger.The paper attributes GAG’s storage efficiency to its stronger performance with the smaller reservoir.Source: §5.3.2 · pp. 8–9, Figure 8 · p. 9
  • Larger window sizes yield better recommendation performance because they let the model use more data for updates.Source: Figure 10 · p. 9, §5.3.2 · p. 9

5.4 Parameter Sensitivity

The experiments evaluate GAG’s sensitivity to embedding size and model depth. Performance is highest with embedding size 200 and declines as the number of layers increases.

5.4 Parameter Sensitivity

  • The section evaluates parameter sensitivity in the GAG model.Source: §5.4 · p. 9

5.4.1 Embedding Size.

  • Embedding size 200 gives GAG the highest performance in all reported situations; size 100 is relatively strong compared with 50 and 400.The authors characterize sizes 50 and 400 as respectively unrepresentative and over-parameterized for their methods, making strong-model training difficult.Source: §5.4.1 · p. 9

5.4.2 Number of Layers.

  • GAG performance decreases as the model becomes deeper across the tested one-, two-, and three-layer variants.The authors also note that session connectivity is smaller than in traditional graph data, limiting deeper GNN models.Source: §5.4.2 · p. 9

6 CONCLUSION

The paper concludes that GAG uses a global attribute to address preserving users’ long-term interests in SSR, alongside a Wasserstein reservoir for session sampling. It identifies incorporating cross-session information as future work.

6 CONCLUSION

  • GAG addresses preserving users’ long-term interests in SSR through the global attribute and GAG layer.Source: §6 · p. 9
  • The Wasserstein reservoir samples sessions according to the distance between recommendation results and real interactions.Source: §6 · p. 10
  • The conclusion identifies incorporating cross-session information as future work for SSR.Source: §6 · p. 10

Related papers

Loading 2007.02747v2…