Overview

This project presents a novel framework combining hypergraphs with large language models (LLMs) to analyze personality traits from online social networks. The project aims to overcome the limitations of traditional data mining methods, and leveraging the associative capabilities of LLMs and the structural potential of hypergraphs to provide a more profound analysis of user behavior and interactions within dynamic social flows and networks in digital realms.

Key Contributions

This project makes three significant contributions to the field:

1. Prompt-based Personality Extraction with LLM: We have designed a novel prompt-based method to effectively extract users' personality traits from large language models.
2. Data Collection and Analysis: We performed extensive data collection and analysis from the Personality Cafe forum, enabling comprehensive insights into user profiles and interactions.
3. Hypergraph neural network for social network simulation: We proposed a new model using Deep hypergraphs to capture the intricate relationships among users and their personality traits. This model can be used to depict social environments and energy flows in real-world scenarios.

Datasets

We collected totally 85462 users profiles from Personality Cafe, with the following information:

• Usernames
• MBTI types
• Gender
• Followers
• Self-descriptions (About section)
• Sexual orientation
• Enneagram Type

To speed up, we selected 17000 users with both completed followers, groups, MBTI and Enneagram information to generate natrual-language descriptions. The dataset is stored in dataset.

Settings

To run the code, simply clone the repository and install the required packages:

git clone https://github.com/ZhiyaoShu/LLM-HGNN-MBTI.git
cd LLM-HGNN-MBTI
pip install -r requirements.txt

Test pre-trained models

You can run the test.py to test a pre-trained hypergraph neural network(HGNN) with following arguments:

python test.py --test_model_path best_model_hgnn.pth 

You can also test the hypergraph neural network plus(HGNNP) and change the test_model_path to best_model_hgnnp.pth

Note that we suppose you download the pre-trained models in the repo root directory.

Training

To train a model, you need to:

- Natrual-language descriptions and converted embeddings.

As many new LLMs emerged after we publish, you can either generate new features with SOTAs with row data, or run with the existed generated descriptions features from the GPT-3.5-turbo, converted by sentence-transformers. You can download the descriptions and features from dataset:

• Generated descriptions

• Converted embeddings

• You can also downloaded processed feature maps, which has aggregated user inforamtion and descriptions.

- Three types hyperedges. You can download structured hyperedges here

After you prepare previous steps, you can start training the model with the following arguments:

python train.py

Check the parser arguments to adjust output path, model types, epoches and other parameters.

Contribution & Collaborations

We encourage the community to contribute to this project. Feel free to send us feedback, suggest improvements, or submit pull requests with your innovative ideas and changes.

Zhiyao Shu Xiangguo Sun

References

DHG OPENAI API LLAMA Google Gemma

🌹Please Cite Our Work If Helpful:

Thanks! / 谢谢! / ありがとう! / merci! / 감사! / Danke! / спасибо! / gracias! ...

@inproceedings{shu2024llm,
  title={When LLM Meets Hypergraph: A Sociological Analysis on Personality via Online Social Networks},
  author={Shu, Zhiyao and Sun, Xiangguo and Cheng, Hong},
  booktitle={Proceedings of the 33th ACM international conference on information \& knowledge management (CIKM)},
  year={2024}
}

Works related to this dataset and graph with social personality analysis:

@article{sun2023self,
  title={Self-supervised hypergraph representation learning for sociological analysis},
  author={Sun, Xiangguo and Cheng, Hong and Liu, Bo and Li, Jia and Chen, Hongyang and Xu, Guandong and Yin, Hongzhi},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  volume={35},
  number={11},
  pages={11860--11871},
  year={2023},
  publisher={IEEE}
}

@article{sun2022your,
  title={In your eyes: Modality disentangling for personality analysis in short video},
  author={Sun, Xiangguo and Liu, Bo and Ai, Liya and Liu, Danni and Meng, Qing and Cao, Jiuxin},
  journal={IEEE Transactions on Computational Social Systems},
  volume={10},
  number={3},
  pages={982--993},
  year={2022},
  publisher={IEEE}
}

@article{sun2020group,
  title={Group-level personality detection based on text generated networks},
  author={Sun, Xiangguo and Liu, Bo and Meng, Qing and Cao, Jiuxin and Luo, Junzhou and Yin, Hongzhi},
  journal={World Wide Web},
  volume={23},
  pages={1887--1906},
  year={2020},
  publisher={Springer}
}

@inproceedings{sun2018personality,
  title={Who am I? Personality detection based on deep learning for texts},
  author={Sun, Xiangguo and Liu, Bo and Cao, Jiuxin and Luo, Junzhou and Shen, Xiaojun},
  booktitle={2018 IEEE international conference on communications (ICC)},
  pages={1--6},
  year={2018},
  organization={IEEE}
}