Research map / Adversarial robustness and security / Certified and empirical robustness
Sparse Enhanced Network: An Adversarial Generation Method for Robust Augmentation in 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 SparseEnNet, an adversarial generation method that creates deep (embedding-space) augmented sequences for sequential recommendation, trained with discriminators, negative-sample learning and clustering-based self-training.
Problem
Sequential recommendation suffers when user histories are short and sparse, and existing shallow sequence augmentations (crop, mask, reorder) do not explore the deep embedding space of sequences and can hurt accuracy when overused.
Approach
Combines shallow augmentations with a learned 'pooling' deep augmentation and trains an encoder adversarially via an augmentation discriminator (gradient reversal) and a stability discriminator, plus InfoNCE negative-sample learning and k-means-based self-training loss, so the encoder learns augmentation-invariant representations.
Key results
- SparseEnNet improves Hit@5 over the second-best model by 11.93% (Beauty), 6.91% (Toys), and 4.27% (Yelp).
- NDCG@5 gains over the second-best model are 14.47% (Beauty), 7.09% (Toys), and 6.21% (Yelp).
- Ablation on Beauty shows removing negative sample learning drops Hit@10 from 0.0762 to 0.0649.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SparseEnNet instead of ICLRec | Sequential recommendation · Beauty | Hit@5: 0.0461 -> 0.0516 | Helped |
| Use SparseEnNet instead of ICLRec | Sequential recommendation · Beauty | NDCG@5: 0.0304 -> 0.0348 | Helped |
| Use SparseEnNet instead of ICLRec | Sequential recommendation · Toys | Hit@5: 0.0579 -> 0.0619 | Helped |
| Use SparseEnNet instead of ICLRec | Sequential recommendation · Yelp | Hit@5: 0.0234 -> 0.0244 | Helped |
| Add stability discriminator (ablation) | SparseEnNet · Beauty and Toys | Hit@10: 0.0737 -> 0.0762; 0.0847 -> 0.0855 | Mixed |
| Add augmentation discriminator (ablation) | SparseEnNet · Beauty and Toys | Hit@10: 0.0738 -> 0.0762; 0.0870 -> 0.0855 | Mixed |
| Add self-training enhanced learning (ablation) | SparseEnNet · Beauty and Toys | Hit@10: 0.0713 -> 0.0762; 0.0805 -> 0.0855 | Helped |
| Add negative sample learning (ablation) | SparseEnNet · Beauty and Toys | Hit@10: 0.0649 -> 0.0762; 0.0775 -> 0.0855 | Helped |
| Add all discriminators (ablation) | SparseEnNet · Beauty and Toys | Hit@10: 0.0724 -> 0.0762; 0.0865 -> 0.0855 | Mixed |
| Add reorder augmentation (ablation) | SparseEnNet · Beauty | Hit@5: 0.0471 -> 0.0516 | Helped |
| Add reorder augmentation (ablation) | SparseEnNet · Beauty | NDCG@5: 0.0304 -> 0.0348 | Helped |
Surprising findings
- Removing the augmentation discriminator slightly improves Hit@10 on Toys (0.0870 vs 0.0855), and removing all discriminators also slightly improves Toys Hit@10 (0.0865), suggesting discriminators do not uniformly help.
Limitations
- Excessive data augmentation can hurt accuracy on longer sequences, as the authors note.
- Cold-start evaluation is a simulated case study on sampled items, and SparseEnNet performs at a similar level to ICLRec on popular items.
- Hyperparameter results are deferred to supplementary material; evaluation limited to three datasets.
Experimental setup
Datasets: Amazon Beauty Amazon Toys Yelp
Benchmarks: Amazon Beauty Amazon Toys Yelp
Metrics: Hit@5 NDCG@5 Hit@10 NDCG@10 Hit@20 NDCG@20
Baselines: GRU4Rec SASRec BERT4Rec CL4SRec ICLRec
Builds on: CL4SRec ICLRec SASRec Gradient reversal layer (domain adversarial training) InfoNCE
Closest papers by meaning
- Iterative Sparse Attention for Long-sequence Recommendation (AAAI 2025 · similarity 0.87)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.86)
- Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.86)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · 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.