Research map / Graph learning / Knowledge and temporal graphs
Spatial-Temporal Knowledge Distillation for Takeaway 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 STKDRec, a two-stage model that pre-trains a GNN teacher on a spatial-temporal knowledge graph and distills its soft predictions into a lightweight spatial-temporal Transformer for takeaway recommendation.
Problem
Takeaway recommendation from purchase sequences suffers from sparsity and fails to capture dynamic user preferences over fine-grained geospatial information, while integrating graph-based spatial-temporal knowledge is computationally expensive.
Approach
Train a GNN-based STKG encoder as a teacher on sampled subgraphs, then transfer its predicted distributions to a sequence-based ST-Transformer (with spatial region and distance embeddings) via KL-based knowledge distillation, avoiding GNN cost at inference.
Key results
- On Wuhan, STKDRec reaches HR@10 0.8229 and NDCG@10 0.7586, versus 0.7976 and 0.7318 for the best baselines (MAERec and BSARec).
- On Sanya, STKDRec reaches HR@10 0.8940 and NDCG@10 0.8492, above the second-best BSARec at 0.8838 and 0.8259.
- Ablation shows removing KD (-w/o KD) drops Wuhan HR@10 from 0.8229 to 0.8161, and removing both spatial position and KD drops it to 0.7796.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add spatial-temporal knowledge distillation to the ST-Transformer (ablation) | STKDRec · Wuhan | HR@10: 0.8161 -> 0.8229 | Helped |
| Add spatial-temporal knowledge distillation to the ST-Transformer (ablation) | STKDRec · Sanya | HR@10: 0.8838 -> 0.8940 | Helped |
| Add spatial-temporal knowledge distillation to the ST-Transformer (ablation) | STKDRec · Taiyuan | HR@10: 0.8684 -> 0.8789 | Helped |
| Add spatial position embedding (ablation) | STKDRec · Wuhan | HR@10: 0.8097 -> 0.8229 | Helped |
| Add spatial position embedding (ablation) | STKDRec · Sanya | HR@10: 0.8844 -> 0.8940 | Helped |
| Add spatial position embedding (ablation) | STKDRec · Taiyuan | HR@10: 0.8669 -> 0.8789 | Helped |
| Add spatial region information alongside spatial distance (ablation) | STKDRec · Wuhan | HR@10: 0.8181 -> 0.8229 | Helped |
| Add spatial region information alongside spatial distance (ablation) | STKDRec · Sanya | HR@10: 0.8875 -> 0.8940 | Helped |
| Add spatial region information alongside spatial distance (ablation) | STKDRec · Taiyuan | HR@10: 0.8685 -> 0.8789 | Helped |
| Add spatial distance information alongside spatial regions (ablation) | STKDRec · Wuhan | HR@10: 0.8178 -> 0.8229 | Helped |
| Add spatial distance information alongside spatial regions (ablation) | STKDRec · Sanya | HR@10: 0.8877 -> 0.8940 | Helped |
| Add spatial distance information alongside spatial regions (ablation) | STKDRec · Taiyuan | HR@10: 0.8688 -> 0.8789 | Helped |
Limitations
- Evaluation uses 100 randomly sampled negatives, which may inflate absolute metrics compared with full ranking.
- Efficiency claims are not quantified in the main text.
- Evaluated only on three Ele.me city datasets.
Experimental setup
Datasets: Wuhan takeaway dataset (Ele.me) Sanya takeaway dataset (Ele.me) Taiyuan takeaway dataset (Ele.me)
Benchmarks: Wuhan Sanya Taiyuan
Metrics: HR@5 HR@10 HR@20 NDCG@5 NDCG@10 NDCG@20
Baselines: Caser GRU4Rec SASRec BERT4Rec DuoRec FEARec GCL4SR MAERec BSARec
Builds on: knowledge distillation SASRec TinyBERT
Closest papers by meaning
- GeoMamba: Towards Multi-granular POI Recommendation with Geographical State Space Model (AAAI 2025 · similarity 0.85)
- SPOT-Trip: Dual-Preference Driven Out-of-Town Trip Recommendation (NeurIPS 2025 · similarity 0.85)
- Integrating Personalized Spatio-Temporal Clustering for Next POI Recommendation (AAAI 2025 · similarity 0.84)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.84)
- TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation (NeurIPS 2025 · similarity 0.84)
- Enhancing Long-and Short-Term Representations for Next POI Recommendations via Frequency and Hierarchical Contrastive Learning (AAAI 2025 · similarity 0.83)
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.