LLMRec: Large Language Models with Graph Augmentation for Recommendation

PyTorch implementation for WSDM 2024 paper LLMRec: Large Language Models with Graph Augmentation for Recommendation.

Wei Wei, Xubin Ren, Jiabin Tang, Qingyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang, Dawei Yin and Chao Huang*. (*Correspondence)

Data Intelligence Lab@University of Hong Kong, Baidu Inc.

This repository hosts the code, original data and augmented data of LLMRec.

LLMRec is a novel framework that enhances recommenders by applying three simple yet effective LLM-based graph augmentation strategies to recommendation system. LLMRec is to make the most of the content within online platforms (e.g., Netflix, MovieLens) to augment interaction graph by i) reinforcing u-i interactive edges, ii) enhancing item node attributes, and iii) conducting user node profiling, intuitively from the natural language perspective.

🎉 News 📢📢

[2024.3.20] 🚀🚀 📢📢📢📢🌹🔥🔥🚀🚀 Because baselines LATTICE and MMSSL require some minor modifications, we provide code that can be easily run by simply modifying the dataset path.
[2023.11.3] 🚀🚀 Release the script for constructing the prompt.
[2023.11.1] 🔥🔥 Release the multi-modal datasets (Netflix, MovieLens), including textual data and visual data.
[2023.11.1] 🚀🚀 Release LLM-augmented textual data(by gpt-3.5-turbo-0613), and LLM-augmented embedding(by text-embedding-ada-002).
[2023.10.28] 🔥🔥 The full paper of our LLMRec is available at LLMRec: Large Language Models with Graph Augmentation for Recommendation.
[2023.10.28] 🚀🚀 Release the code of LLMRec.

👉 TODO

Provide different larger version of the datasets.
...

Dependencies

pip install -r requirements.txt

Usage

Stage 1: LLM-based Data Augmentation

cd LLMRec/LLM_augmentation/
python ./gpt_ui_aug.py
python ./gpt_user_profiling.py
python ./gpt_i_attribute_generate_aug.py

Stage 2: Recommender training with LLM-augmented Data

cd LLMRec/
python ./main.py --dataset {DATASET}

Supported datasets: netflix, movielens

Specific code execution example on 'netflix':

# LLMRec
python ./main.py

# w/o-u-i
python ./main.py --aug_sample_rate=0.0

# w/o-u
python ./main.py --user_cat_rate=0

# w/o-u&i
python ./main.py --user_cat_rate=0  --item_cat_rate=0

# w/o-prune
python ./main.py --prune_loss_drop_rate=0

Datasets

├─ LLMRec/ 
    ├── data/
      ├── netflix/
      ...

Multi-modal Datasets

🌹🌹 Please cite our paper if you use the 'netflix' dataset~ ❤️

We collected a multi-modal dataset using the original Netflix Prize Data released on the Kaggle website. The data format is directly compatible with state-of-the-art multi-modal recommendation models like LLMRec, MMSSL, LATTICE, MICRO, and others, without requiring any additional data preprocessing.

Textual Modality: We have released the item information curated from the original dataset in the "item_attribute.csv" file. Additionally, we have incorporated textual information enhanced by LLM into the "augmented_item_attribute_agg.csv" file. (The following three images represent (1) information about Netflix as described on the Kaggle website, (2) textual information from the original Netflix Prize Data, and (3) textual information augmented by LLMs.)

Visual Modality: We have released the visual information obtained from web crawling in the "Netflix_Posters" folder. (The following image displays the poster acquired by web crawling using item information from the Netflix Prize Data.)

Original Multi-modal Datasets & Augmented Datasets

Download the Netflix dataset.

🚀🚀 We provide the processed data (i.e., CF training data & basic user-item interactions, original multi-modal data including images and text of items, encoded visual/textual features and LLM-augmented text/embeddings). 🌹 We hope to contribute to our community and facilitate your research 🚀🚀 ~

netflix: Google Drive Netflix. 🌟(Image&Text)

Encoding the Multi-modal Content.

We use CLIP-ViT and Sentence-BERT separately as encoders for visual side information and textual side information.

Prompt & Completion Example

LLM-based Implicit Feedback Augmentation

Prompt

Recommend user with movies based on user history that each movie with title, year, genre. History: [332] Heart and Souls (1993), Comedy|Fantasy [364] Men with Brooms(2002), Comedy|Drama|Romance Candidate: [121]The Vampire Lovers (1970), Horror [155] Billabong Odyssey (2003),Documentary [248]The Invisible Guest 2016, Crime, Drama, Mystery Output index of user's favorite and dislike movie from candidate.Please just give the index in [].

Completion

248 121

LLM-based User Profile Augmentation

Prompt

Generate user profile based on the history of user, that each movie with title, year, genre. History: [332] Heart and Souls (1993), Comedy|Fantasy [364] Men with Brooms (2002), Comedy|Drama|Romance Please output the following infomation of user, output format: {age: , gender: , liked genre: , disliked genre: , liked directors: , country: , language: }

Completion

{age: 50, gender: female, liked genre: Comedy|Fantasy, Comedy|Drama|Romance, disliked genre: Thriller, Horror, liked directors: Ron Underwood, country: Canada, United States, language: English}

LLM-based Item Attributes Augmentation

Prompt

Provide the inquired information of the given movie. [332] Heart and Souls (1993), Comedy|Fantasy The inquired information is: director, country, language. And please output them in form of: director, country, language

Completion

Ron Underwood, USA, English

Augmented Data

Augmented Implicit Feedback (Edge)

For each user, 0 represents a positive sample, and 1 represents a negative sample.

Augmented User Profile (User Node)

For each user, the dictionary stores augmented information such as 'age,' 'gender,' 'liked genre,' 'disliked genre,' 'liked directors,' 'country,' and 'language.'

Augmented item attribute

For each item, the dictionary stores augmented information such as 'director,' 'country,' and 'language.'

Candidate Preparing for LLM-based Implicit Feedback Augmentation

step 1: select base model such as MMSSL or LATTICE

step 2: obtain user embedding and item embedding

step 3: generate candidate

      _, candidate_indices = torch.topk(torch.mm(G_ua_embeddings, G_ia_embeddings.T), k=10)  
      pickle.dump(candidate_indices.cpu(), open('./data/' + args.datasets +  '/candidate_indices','wb'))

Example of specific candidate data.

In [3]: candidate_indices
Out[3]: 
tensor([[ 9765,  2930,  6646,  ..., 11513, 12747, 13503],
        [ 3665,  8999,  2587,  ...,  1559,  2975,  3759],
        [ 2266,  8999,  1559,  ...,  8639,   465,  8287],
        ...,
        [11905, 10195,  8063,  ..., 12945, 12568, 10428],
        [ 9063,  6736,  6938,  ...,  5526, 12747, 11110],
        [ 9584,  4163,  4154,  ...,  2266,   543,  7610]])

In [4]: candidate_indices.shape
Out[4]: torch.Size([13187, 10])

Citing

If you find this work helpful to your research, please kindly consider citing our paper.

@article{wei2023llmrec,
  title={LLMRec: Large Language Models with Graph Augmentation for Recommendation},
  author={Wei, Wei and Ren, Xubin and Tang, Jiabin and Wang, Qinyong and Su, Lixin and Cheng, Suqi and Wang, Junfeng and Yin, Dawei and Huang, Chao},
  journal={arXiv preprint arXiv:2311.00423},
  year={2023}
}

Acknowledgement

The structure of this code is largely based on MMSSL, LATTICE, MICRO. Thank them for their work.

hkuds / llmrec Goto Github PK

llmrec's Introduction