Research map / Graph learning / Knowledge and temporal graphs
Temporal Graph Contrastive Learning for Sequential 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 TGCL4SR, a sequential recommender that combines a temporal item transition graph (with neighbor-sampling and time-disturbance augmentations, a TiTConv GNN, and graph contrastive losses) with a time-interval-aware Transformer, trained with MMD auxiliary loss.
Problem
Sequential recommendation must exploit multi-type temporal information and cope with noise and sparsity in evolving user behaviors, which existing sequence and graph models handle poorly.
Approach
Builds a global Temporal Item Transition Graph whose edges carry timestamps and user attributes, augments it via neighbor sampling and Gaussian time disturbance, encodes it with a time-aware graph convolution (TiTConv), and aligns views with subgraph and disturbed graph contrastive losses, while a Transformer with time-interval position embeddings models local sequences.
Key results
- TGCL4SR improves over the best baseline by 1.08% to 12.4% across datasets on HR and NDCG metrics (Table 2).
- Ablation on Beauty and Comics shows removing the temporal graph contrastive learning loss hurts most among components.
- Gains are smaller on CDs than on Beauty, Games and Comics, attributed to longer sequences in CDs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TGCL4SR instead of the strongest baseline, GCL4SR | Sequential recommendation · Beauty | HR@10, HR@20, NDCG@10, NDCG@20 | Helped |
| Use TGCL4SR instead of the strongest baseline, GCL4SR | Sequential recommendation · Games | HR@10, HR@20, NDCG@10, NDCG@20 | Helped |
| Use TGCL4SR instead of the strongest baseline, GCL4SR | Sequential recommendation · CDs | HR@10, HR@20, NDCG@10, NDCG@20 | Mixed |
| Use TGCL4SR instead of the strongest baseline, GCL4SR | Sequential recommendation · Comics | HR@10, HR@20, NDCG@10, NDCG@20 | Helped |
| Add disturbed graph contrastive learning (DGCL) (ablation) | TGCL4SR · Beauty and Comics | NDCG@10, NDCG@20 | Helped |
| Add subgraph contrastive learning (SGCL) (ablation) | TGCL4SR · Beauty and Comics | NDCG@10, NDCG@20 | Helped |
| Add temporal graph contrastive learning (TGCL) (ablation) | TGCL4SR · Beauty and Comics | NDCG@10, NDCG@20 | Helped |
| Use TiTConv instead of stacked GraphSAGE (ablation) | TGCL4SR · Beauty and Comics | NDCG@10, NDCG@20 | Helped |
| Add time-interval embeddings to the sequence encoder (ablation) | TGCL4SR · Beauty and Comics | NDCG@10, NDCG@20 | Helped |
Surprising findings
- Gains are smaller on the CDs dataset, which the authors attribute to larger sequence scale alleviating transition-pattern difficulty.
Limitations
- Gains on the CDs dataset are limited relative to other datasets.
- Temporal graph is large and multi-edge, so computing representations over the full graph is costly, motivating sampling.
Experimental setup
Datasets: Amazon Beauty Amazon Video Games Amazon CDs Goodreads Comics Graphic
Benchmarks: Amazon Beauty Amazon Video Games Amazon CDs Goodreads Comics Graphic
Metrics: HR@10 HR@20 NDCG@10 NDCG@20
Baselines: GRU4Rec SASRec TiSASRec CL4SRec DuoRec TCPSRec GCL4SR MStein
Builds on: SASRec TiSASRec GraphSAGE InfoNCE GCL4SR
Closest papers by meaning
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.89)
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation (NeurIPS 2025 · similarity 0.88)
- TGB-Seq Benchmark: Challenging Temporal GNNs with Complex Sequential Dynamics (ICLR 2025 · similarity 0.87)
- SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.87)
- LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.87)
- Enhancing Diffusion Model with Auxiliary Information Mining-Exploration and Efficient Sampling Mechanism for Sequential Recommendation (AAAI 2025 · similarity 0.86)
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.