Paper: arXiv 2503.14213

Authors: Ashraf Ghiye, Baptiste Barreau, Laurent Carlier, Michalis Vazirgiannis

Abstract

Graph Neural Networks have significantly advanced research in recommender systems over the past few years. These methods typically capture global interests using aggregated past interactions and rely on static embeddings of users and items over extended periods of time. While effective in some domains, these methods fall short in many real-world scenarios, especially in finance, where user interests and item popularity evolve rapidly over time. To address these challenges, we introduce a novel extension to Light Graph Convolutional Network (LightGCN) designed to learn temporal node embeddings that capture dynamic interests. Our approach employs causal convolution to maintain a forward-looking model architecture. By preserving the chronological order of user-item interactions and introducing a dynamic update mechanism for embeddings through a sliding window, the proposed model generates well-timed and contextually relevant recommendations. Extensive experiments on a real-world dataset from BNP Paribas demonstrate that our approach significantly enhances the performance of LightGCN while maintaining the simplicity and efficiency of its architecture. Our findings provide new insights into designing graph-based recommender systems in time-sensitive applications, particularly for financial product recommendations.

Complexity vs Empirical Score

  • Math Complexity: 5.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper uses standard Graph Neural Network equations and causal convolution formalism (moderate math), and validates the method with extensive experiments on a real-world financial dataset from BNP Paribas, reporting specific performance gains like 4x improvement in mean Average Precision.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Enhance LightGCN for dynamic<br>financial bond recommendations"] --> B["Methodology<br>Integrate causal convolution with<br>LightGCN architecture"]
  B --> C["Data Input<br>BNP Paribas historical<br>user-item interaction dataset"]
  C --> D["Computational Process<br>Sliding window temporal<br>embedding updates"]
  D --> E["Key Finding 1<br>Maintains LightGCN efficiency<br>with time-sensitive modeling"]
  D --> F["Key Finding 2<br>Significant performance<br>improvement in finance domain"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a novel extension to LightGCN that uses causal convolutions to learn temporal node embeddings for dynamic recommendation in finance.

Source sentence
we introduce a novel extension to Light Graph Convolutional Network (LightGCN) designed to learn temporal node embeddings that capture dynamic interests.

Assumes 4

  • Training uses only recent transactions within a sliding window to avoid future data leakage. model
    Source sentence
    we propose constructing the training graph dynamically using only recent transactions within
  • A predetermined window size controls data drift and maintains recommendation quality. model
    Source sentence
    we fix a predetermined window size to control the effect of data drift and maintain the quality of recommendations.
  • All models except baselines are trained with Bayesian Personalized Ranking loss. model
    Source sentence
    All models, except for MostPop and RecentPop, are trained using the Bayesian Personalized Ranking (BPR) loss
  • User preferences are captured by recent interactions, older transactions are less relevant. data
    Source sentence
    like finance, users who interact with the same items within a shorter time period show stronger similarities.

Data

  • BNP Paribas credit bonds dataset · BNP Paribas alternative proprietary
    01/01/2019 to 01/06/2023 · over 5,000 unique clients and 47,000 unique bonds
    Source sentence
    We run our experiments on a proprietary dataset provided by BNP Paribas, comprising more than 7 million daily transactions spanning a period of five and a half years

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • LightGCN-W with a window size of 2 improves mean Average Precision by 244% (from 1.47 to 5.06) compared to standard LightGCN. performance
    Source sentence
    2, significantly outperforms the standard LightGCN, increasing mAP by 244% from 1.47 to 5.06.
  • The proposed method achieves up to 4× improvement in precision over the conventional training method without compromising LightGCN's simplicity. performance
    Source sentence
    Precision by up to 4x over the conventional training method, without compromising the simplicity and lightweight nature of LightGCN.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 19 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (40)

  1. Temporal graph models fail to capture global temporal dynamics (2023) — link
  2. Towards Better Evaluation for Dynamic Link Prediction (2022) — link
  3. Using collaborative filtering to weave an information tapestry (1992) — link
  4. Graph Neural Networks in Recommender Systems: A Survey (2022) — link
  5. Time Weight Content-Based Extensions of Temporal Graphs for Personalized Recommendation (2017) — link
  6. A Deep Temporal Collaborative Filtering Recommendation Framework via Joint Learning from Long and Short-Term Effects (2019) — link
  7. A Temporal Graph Network Framework for Dynamic Recommendation (2024) — link
  8. Neural Collaborative Filtering (2017) — link
and 32 more
  1. Temporal Collaborative Filtering with Bayesian Probabilistic Tensor Factorization (2010) — link
  2. Deep Time-Aware Matrix Factorization (2020) — link
  3. Collaborative filtering with temporal dynamics (2009) — link
  4. Time-aware recommender systems: a comprehensive survey and analysis of existing evaluation protocols (2014) — link
  5. Time-Aware Recommender Systems: A Systematic Mapping (2017) — link
  6. A Critical Study on Data Leakage in Recommender System Offline Evaluation (2023) — link
  7. Graph Convolutional Neural Networks for Web-Scale Recommender Systems (2018) — link
  8. Semi-Supervised Classification with Graph Convolutional Networks (2017)
  9. Matrix Factorization Techniques for Recommender Systems (2009) — link
  10. A Survey of Graph Neural Networks for Recommender Systems: Challenges, Methods, and Directions (2023) — link
  11. Information Retrieval: Implementing and Evaluating Search Engines (2010)
  12. Item-based top-N recommendation algorithms (2004) — link
  13. Adam: A Method for Stochastic Optimization (2014)
  14. Performance of Recommendation Systems in Dynamic Streaming Environments (2007) — link
  15. Time weight collaborative filtering (2005) — link
  16. Take a Fresh Look at Recommender Systems from an Evaluation Standpoint (2023) — link
  17. Considering temporal aspects in recommender systems: a survey (2023) — link
  18. Dynamic Graph Collaborative Filtering (2020) — link
  19. HOP-rec: high-order proximity for implicit recommendation (2018) — link
  20. Challenging the Myth of Graph Collaborative Filtering: a Reasoned and Reproducibility-driven Analysis (2023) — link
  21. Inductive representation learning on temporal graphs (2020) — link
  22. Temporal Graph Networks for Deep Learning on Dynamic Graphs (2020)
  23. ROLAND: Graph Learning Framework for Dynamic Graphs (2022) — link
  24. A Re -visit of the Popularity Baseline in Recommender Systems (2020) — link
  25. Evaluating Recommendation Systems (2011) — link
  26. Optimizing top-n collaborative filtering via dynamic negative item sampling (2013) — link
  27. BPR: Bayesian personalized ranking from implicit feedback (2009)
  28. Are We Forgetting Something? Correctly Evaluate a Recommender System With an Optimal Training Window. (2022)
  29. LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation (2020) — link
  30. Neural Graph Collaborative Filtering (2019) — link
  31. History-Augmented Collaborative Filtering for Financial Recommendations (2020) — link
  32. Adaptive Collaborative Filtering with Personalized Time Decay Functions for Financial Product Recommendation — → 2308.01208

1 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.