| # MARS v3: Beating SASRec on Sequential Recommendation | |
| ## Results on MovieLens-1M (Full Ranking, 3416 items) | |
| | Model | HR@5 | HR@10 | HR@20 | NDCG@10 | MRR@10 | | |
| |-------|------|-------|-------|---------|--------| | |
| | SASRec (CE loss) | 0.0480 | 0.0826 | 0.1144 | 0.0385 | 0.0251 | | |
| | **MARS-cl02-f3** | 0.0538 | 0.0854 | 0.1197 | 0.0390 | 0.0248 | | |
| | **MARS-cl005-f2** | 0.0507 | 0.0829 | 0.1149 | 0.0383 | 0.0248 | | |
| | **MARS-cl01-f2-d15** | 0.0507 | 0.0826 | 0.1146 | 0.0382 | 0.0246 | | |
| ## Architecture | |
| - Long-term: FMLP FFT filters (O(n log n)) + Compressive Memory | |
| - Short-term: Causal Self-Attention | |
| - Training: Full Softmax CE + DuoRec Dropout Contrastive (InfoNCE) | |
| - Adaptive per-user fusion gate | |