CoRRe

Official github repository for Training-Free LLM-Based Recommendation with Post-LLM Item Refinement Using Collaborative Signals. (Published as a conference paper at CIKM 2026)

• Online Appendix: See [Online Appendix] CoRRe.pdf in the root folder.

This repo follows a 3-step pipeline:

• 🧹 refinement.py: Propagate GPT item features with co-purchase graph (i.e., direction refinement) and magnitude refinement.
• 🧠 generate_user_profile.py: Generate user preference sentences from history.
• 🔎 item_retrieval.py: Retrieve candidates from propagated item embeddings and run Hit@K / NDCG evaluation.

All scripts now use dataset paths under data/{dataset} by default. (sports, toys, beauty)

• Set your OpenAI API key

export OPENAI_API_KEY=your-openai-api-key

• Expected input layout

data/{dataset}/
├── inter.json: User-item interaction data.
├── item2id.json: Mapping from item IDs to internal indices.
├── user2id.json: Mapping from user IDs to internal indices.
├── title.pickle: Item titles.
├── gpt_feat.npy: LLM-based embeddings of item titles.
├── generated_intent.json: LLM-generated user intents/profiles.
└── refined_itemfeat.npy: Refined item embeddings.

{dataset} should be one of: sports, toys, beauty

Due to file size limitations, gpt_feat.npy and refined_itemfeat.npy are provided separately. Please download them from the following Dropbox link and place them in the appropriate feature directory: https://www.dropbox.com/scl/fo/rtloyxm3qoawuclb2utgf/AOrrzXzDQN7hYT154RXY8jU?rlkey=5jprf6flw3g9x7hxcqj1vkzl2&st=uy53dm1v&dl=0

[Sports] python refinement.py --dataset sports --fusion_alpha 0.4 --degree_alpha 0.1
[Toys] python refinement.py --dataset toys --fusion_alpha 0.8 --degree_alpha 0.05
[Beauty] python refinement.py --dataset beauty --fusion_alpha 0.2 --degree_alpha 0.05

• data/{dataset}/refined_itemfeat.npy (You can use our uploaded version.)

[Sports] python generate_user_profile.py --dataset sports
[Toys] python generate_user_profile.py --dataset toys
[Beauty] python generate_user_profile.py --dataset beauty

• data/{dataset}/generated_intent.json (You can use our uploaded version.)

Without using an LLM, you can simply average the title embeddings of the user's historical items to obtain a query embedding.

[Sports] python item_retrieval.py --dataset sports --topk 10 20 30 50
[Toys] python item_retrieval.py --dataset toys --topk 10 20 30 50
[Beauty] python item_retrieval.py --dataset beauty --topk 10 20 30 50

• refinement.py

  • --dataset : sports (default), toys, beauty
  • --num_layers : number of propagation layers (default: 1)
  • --fusion_alpha : combined ratio between original and propagated embeddings
  • --apply_degree_multiply / --no-apply_degree_multiply : multiply each final item embedding by d^degree_alpha (default: enabled)
  • --degree_alpha : exponent for degree scaling (default: 0.0)

• generate_user_profile.py

  • --dataset : sports (default), toys, beauty
  • --max_users : number of users to process (None for all), default is 1,000 in our paper.
  • --max_items_per_user : number of most recent items from history (default: 10)
  • --model : chat model name

• item_retrieval.py

  • --dataset : sports (default), toys, beauty
  • --topk : list of Top-K values
  • --model : embedding model
  • --degree_alpha : popularity re-ranking factor
  • --mask_train_items / --no_mask_train_items : whether to exclude train items from retrieval candidates