Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Evidential Stochastic Differential Equations for Time-Aware 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 E-NSDE, which combines neural stochastic differential equations for continuous user/item representations with evidential deep learning and a monotonic network so that uncertainty grows with the interaction time gap, for time-aware sequential recommendation.
Problem
Sequential recommenders assume uniform time intervals between interactions and ignore that uncertainty in user preference grows with the gap since the last interaction, so they capture continuous evolving behavior poorly and cannot quantify uncertainty.
Approach
Integrates an NSDE (drift plus Brownian diffusion) for continuous-time user and item representations with an evidential (NIG) regression head, and uses a monotonic network conditioned on the time gap so that epistemic uncertainty provably increases with the interval; adds an uncertainty-aware weighted BPR loss for exploration.
Key results
- On MovieLens-100K, E-NSDE achieves P@5 0.4711 and nDCG@5 0.4112, versus 0.4398 / 0.3902 for the strongest ODE baseline GRU-ODE.
- Ablation on MovieLens-1M-level results shows the full model (NSDE+EDL+WBPR) reaches P@5 0.4711 vs 0.4065 for NSDE alone.
- Diversity on MovieLens-1M: E-NSDE Gini 0.7413 vs 0.6845 for BERT4Rec, with higher coverage and novelty.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use E-NSDE instead of the strongest reported baseline | Sequential recommendation models · MovieLens-100K | P@5: 0.4398±0.014 -> 0.4711±0.015 | Helped |
| Use E-NSDE instead of the strongest reported baseline | Sequential recommendation models · MovieLens-1M | P@5: 0.4286±0.014 -> 0.4551±0.011 | Helped |
| Use E-NSDE instead of the strongest reported baseline | Sequential recommendation models · Netflix | P@5: 0.3994±0.013 -> 0.4194±0.013 | Helped |
| Use E-NSDE instead of the strongest reported baseline | Sequential recommendation models · Amazon Book | P@5: 0.3884±0.013 -> 0.4021±0.014 | Helped |
| Add EDL to NSDE without WBPR (ablation) | NSDE · MovieLens-100K | P@5: 0.4065 -> 0.4523 | Helped |
| Add EDL to NSDE without WBPR (ablation) | NSDE · MovieLens-100K | nDCG@5: 0.3677 -> 0.3962 | Helped |
| Add WBPR to NSDE without EDL (ablation) | NSDE · MovieLens-100K | P@5: 0.4065 -> 0.4120 | Helped |
| Add WBPR to NSDE without EDL (ablation) | NSDE · MovieLens-100K | nDCG@5: 0.3677 -> 0.3715 | Helped |
| Add EDL and WBPR to NSDE (ablation) | NSDE · MovieLens-100K | P@5: 0.4065 -> 0.4711 | Helped |
| Add EDL and WBPR to NSDE (ablation) | NSDE · MovieLens-100K | nDCG@5: 0.3677 -> 0.4112 | Helped |
| Set evidential regularizer λ to 0.001 (ablation) | E-NSDE · MovieLens-1M | P@5: 0.4236 -> 0.4551 | Helped |
| Set evidential regularizer λ to 0.001 (ablation) | E-NSDE · Netflix | nDCG@5: 0.3574 -> 0.3637 | Helped |
Limitations
- Benefit of time-aware uncertainty is less pronounced when interactions occur at uniform intervals, such as click-through data.
- Evaluated only on rating-based datasets with next-item prediction and a single ground-truth target per sequence.
Experimental setup
Datasets: MovieLens-100K MovieLens-1M Netflix Amazon Book BookCrossing Yahoo Music
Benchmarks: MovieLens-100K MovieLens-1M Netflix Amazon Book BookCrossing Yahoo Music
Metrics: P@5 nDCG@5 P@10 P@20 nDCG@10 nDCG@20 Gini index coverage novelty
Baselines: timeSVD++ CKF NGCF LightGCN CASER SASRec BERT4Rec S3-Rec CL4SRec SAR ResAct LT-OCF
Builds on: neural ODE evidential deep learning BERT4Rec GRU-ODE BPR
Closest papers by meaning
- LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.85)
- Influential Exemplar Replay for Incremental Learning in Recommender Systems (AAAI 2024 · similarity 0.84)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.84)
- Breaking Determinism: Fuzzy Modeling of Sequential Recommendation Using Discrete State Space Diffusion Model (NeurIPS 2024 · similarity 0.84)
- SIGMA: Selective Gated Mamba for Sequential Recommendation (AAAI 2025 · similarity 0.84)
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.