Research map / Reinforcement learning / Imitation learning
POI Recommendation via Multi-Objective Adversarial Imitation Learning
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 MOAIR, a generative adversarial imitation learning recommender for next-POI prediction that uses a graph-enhanced self-supervised state encoder, a variational bottleneck policy, and multi-reward 3PO optimization to stay robust under sparse, noisy check-in data.
Problem
Next POI recommendation from historical check-ins suffers from data sparsity, incompleteness and noise, which degrade deep learning recommenders trained on these trajectories.
Approach
Casts POI recommendation as sequential decision making solved by GAIL-style adversarial imitation, with a state encoder built from Nifty GCNs and spatial-temporal-aware attention, a latent-variable policy regularized by a variational information bottleneck, and multiple rule-based rewards optimized jointly by a parallel PPO (3PO) scheme with gradient aggregation.
Key results
- MOAIR reports Acc@5 of 0.415 on TKY, 0.436 on NYC and 0.225 on Gowalla, versus 0.381, 0.401 and 0.191 for the strongest baseline Graph-Flashback.
- Ablations show removing multi-reward costs about 17% Acc@5 on TKY, removing behavior cloning about 9%, and removing st-attention about 10%.
- On NYC with reduced training data, MOAIR is reported as the least affected of the compared models, though the paper gives only a figure without exact numbers.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MOAIR instead of Graph-Flashback | POI recommendation · Tokyo | Acc@5: 0.381 -> 0.415 | Helped |
| Use MOAIR instead of Graph-Flashback | POI recommendation · New York | Acc@5: 0.401 -> 0.436 | Helped |
| Use MOAIR instead of Graph-Flashback | POI recommendation · Gowalla | Acc@5: 0.191 -> 0.225 | Helped |
| Add multi-reward learning (ablation) | MOAIR · Tokyo | Acc@5: 17% performance drop (Acc@5 in TKY) when removed | Helped |
| Add behavior cloning pretraining (ablation) | MOAIR · Tokyo | Acc@5: 9% performance drop when removed | Helped |
| Use a latent-variable generative policy (ablation) | MOAIR · Tokyo | Acc@5: 5% performance drop when removed | Helped |
| Use spatial-temporal-aware attention instead of distance and time interval embeddings (ablation) | MOAIR · Tokyo | Acc@5: 10% performance drop when removed | Helped |
| Reduce training data to 90%, 75%, 60%, 45%, and 30% of the original data | MOAIR · NYC | Acc@5 | Mixed |
Limitations
- Evaluated only on three check-in datasets with fixed splits and rule-based auxiliary rewards.
- Several components (multi-reward, BC pretraining, latent variable) add complexity; no statistical significance or variance reported in the text.
Experimental setup
Datasets: Foursquare Tokyo Foursquare New York Gowalla
Benchmarks: Foursquare TKY Foursquare NYC Gowalla
Metrics: Acc@5 Acc@10 Prec@5 Prec@10 MRR
Baselines: LSTM ST-RNN HST-LSTM STAN AGRAN GETNext Graph-Flashback
Builds on: GAIL VAIL PPO GETNext STAN
Closest papers by meaning
- Diffusing States and Matching Scores: A New Framework for Imitation Learning (ICLR 2025 · similarity 0.84)
- Exploring Gradient Explosion in Generative Adversarial Imitation Learning: A Probabilistic Perspective (AAAI 2024 · similarity 0.83)
- Adversarial Imitation Learning via Boosting (ICLR 2024 · similarity 0.83)
- Diffusion Imitation from Observation (NeurIPS 2024 · similarity 0.83)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.83)
- Mimicking Better by Matching the Approximate Action Distribution (ICML 2024 · similarity 0.82)
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.