HyperBandit

This repository provides the official PyTorch implementation for the CIKM 23 Oral paper "HyperBandit: Contextual Bandit with Hypernetwork for Time-Varying User Preferences in Streaming Recommendation" and the TOIS paper (HyperBandit+) "Enhancing Bandit Algorithms with LLMs for Time-varying User Preferences in Streaming Recommendations"

Quick Start

Settings

--warm_start

• Type: flag (boolean switch)
• Usage: presence = True, absence = False
• Example: --warm_start
• Description: Enable warm-start training (e.g., initialize from a previous state / pretrained model).
• ✅ HyperBandit: set to False → do not add --warm_start.
• 🟦 HyperBandit+: set to True → add --warm_start.

---

--time_embedding

• Type: string
• Example: --time_embedding="glove"
• Description: Time embedding strategy used to encode temporal information.
• Options: polar / onehot / learn / glove
• ✅ HyperBandit: use --time_embedding="glove".
• 🟦 HyperBandit+: use --time_embedding="polar"

---

--dataset

• Type: string
• Example: --dataset="NYC"
• Description: Dataset identifier; used to select dataset-specific preprocessing and evaluation protocol.
• Options: NYC / TKY / kuai

---

--feature

• Type: string
• Example: --feature="LLM_with_attribute_pca" or --feature="glove_pca"
• Description: Observed feature configuration used as the bandit context representation.
• Options: LLM_with_attribute_pca / glove_pca
• ✅ HyperBandit: use --feature="glove_pca".
• 🟦 HyperBandit+: use --feature="LLM_with_attribute_pca".

bash scripts/quick_start.sh

Citation

If you find our code or paper helpful, please cite our work:

@inproceedings{shen2023hyperbandit,
  title={Hyperbandit: Contextual bandit with hypernewtork for time-varying user preferences in streaming recommendation},
  author={Shen, Chenglei and Zhang, Xiao and Wei, Wei and Xu, Jun},
  booktitle={Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
  pages={2239--2248},
  year={2023}
}