Research map / Graph learning / Knowledge and temporal graphs
LS-TGNN: Long and Short-Term Temporal Graph Neural Network for Session-Based Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LS-TGNN, a temporal GNN for session-based recommendation that models long- and short-term user interests with item-granularity self-supervised disentanglement via triplet loss.
Problem
Session-based recommendation with GNNs ignores the temporal evolution of neighbor relations and of long- versus short-term user intent within sessions.
Approach
Builds a temporal session graph whose neighbor sequences are encoded by a GRU-like time-aware module with item-aware attention, paired with separate long-term (Transformer) and short-term (TGNN) encoders disentangled at item granularity by triplet-loss contrastive proxies.
Key results
- On Tmall, LS-TGNN reaches P@20 of 42.61 vs. 39.28 for the best baseline SPARE, an 8.47% relative improvement.
- On Diginetica, LS-TGNN achieves P@20 55.10 and MRR@20 19.47, versus 54.58/28.09-range baselines such as GCE-GNN (P@20 54.22).
- Ablation shows removing the Short-term Interest Encoder causes the largest performance drop across datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LS-TGNN instead of GCE-GNN | Session-based recommendation · Diginetica | P@20: 54.22 -> 55.10 | Helped |
| Use LS-TGNN instead of COTREC | Session-based recommendation · Diginetica | MRR@20: 19.07 -> 19.47 | Helped |
| Use LS-TGNN instead of SPARE | Session-based recommendation · Retailrocket | P@20: 56.91 -> 58.01 | Helped |
| Use LS-TGNN instead of SPARE | Session-based recommendation · Retailrocket | MRR@20: 30.22 -> 30.74 | Helped |
| Use LS-TGNN instead of SPARE | Session-based recommendation · Tmall | P@20: 39.28 -> 42.61 | Helped |
| Use LS-TGNN instead of SPARE | Session-based recommendation · Tmall | MRR@20: 20.07 -> 20.33 | Helped |
| Add the Long-term Interest Encoder (ablation) | LS-TGNN · Diginetica, Retailrocket, and Tmall | P@20 | Helped |
| Add the Short-term Interest Encoder and temporal session graph (ablation) | LS-TGNN · Diginetica, Retailrocket, and Tmall | P@20 | Helped |
| Add the self-supervised disentangling loss (ablation) | LS-TGNN · Diginetica, Retailrocket, and Tmall | P@20 | Helped |
Limitations
- Parameter sensitivity shows performance degrades when neighbor counts are too large or too small, requiring tuning.
- Temporal GNN approaches are noted as difficult for short sessions with short time intervals, which the paper itself identifies for prior methods.
Experimental setup
Datasets: Tmall RetailRocket Diginetica
Benchmarks: Tmall RetailRocket Diginetica
Metrics: Precision@20 MRR@20
Baselines: GRU4REC NARM STAMP SR-GNN GCE-GNN S2-DHCN COTREC CORE MGIR SPARE
Builds on: SR-GNN GCE-GNN Disentangling long and short-term interests for recommendation CORE
Closest papers by meaning
- Dynamic Multi-Interest Graph Neural Network for Session-Based Recommendation (AAAI 2025 · similarity 0.91)
- ContextGNN: Beyond Two-Tower Recommendation Systems (ICLR 2025 · similarity 0.88)
- Spectral-Based Graph Neural Networks for Complementary Item Recommendation (AAAI 2024 · similarity 0.86)
- Using Time-Aware Graph Neural Networks to Predict Temporal Centralities in Dynamic Graphs (NeurIPS 2024 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.85)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.