life2vec-light

Basic implementation of the life2vec model with the dummy data.

This repository contains basic code for the Using Sequences of Life-events to Predict Human Lives (life2vec) paper. The SocialComplexityLab/life2vec depends (in large) on the specific structure of the data and the version of the packages available.

The code here contains only the model's backbone. Thus, you can easily extend it for your specific use.

A. Simple case

In the simple_workflow.ipynb notebook, we provide code with simple dummy data (using the latest versions of Python packages) to try a simple step-by-step workflow.

Open the simple_workflow.ipynb notebook to try a simple step-by-step workflow.

B. Case with two datasets

In the workflow_with_two_datasets.ipynb, we provide an example for using more than one dataset.

Note: The specification for the environment is located in the requirements.txt (Tutorial on how to set up an environment is here). I have removed the dependency on the conda.

What is missing?

Compared to the original implementation of the life2vec model, this repository does not include the following:

1. Bookkeeping and logging functionality,
2. Data visualisation functionality,
3. Code for experiments, including robustness tests,
4. Implementation of loss and metric functions mentioned in the paper.

Data

You can generate dummy data (that we use here as an example) using the Jupyter Notebook in misc/synthetic_data.ipynb:

1. Generate the dummy user database, aka users.csv,
2. Generate the dummy labor dataset, aka synth_labor.csv,
3. Generate the dummy health dataset, aka synth_health.csv

To Do

• Add code for the pretraining
• Add code with a working data pipeline
• Add a dummy dataset (see labour data in misc/synthetic_data.ipynb)
• Add an example with a noisy data (see health dataset in misc/synthetic_data.ipynb)
• Create a Source file for the Synthetic Labor Data (src/sources/synth_labor.py)
• A working example with two data sources in workflow_two_datasets.ipynb
• Add Logging support
• Add a finetuning example with the specialized decoder

Citations

How to cite THIS code

@misc{https://doi.org/10.5281/zenodo.10680995,
      doi = {10.5281/ZENODO.10680995},
      url = {https://zenodo.org/doi/10.5281/zenodo.10680995},
      author = {Germans Savcisens},
      title = {Github Repository for carlomarxdk/life2vec-light},
      publisher = {Zenodo},
      year = {2024},
      copyright = {Creative Commons Attribution 4.0 International}
}

If you want to cite a specific release, check the DOI number at the top of the README file (or see details of the release).

Connected Materials

@article{savcisens2024using,
      author={Savcisens, Germans and Eliassi-Rad, Tina and Hansen, Lars Kai and Mortensen, Laust Hvas and Lilleholt, Lau and Rogers, Anna and Zettler, Ingo and Lehmann, Sune},
      title={Using sequences of life-events to predict human lives},
      journal={Nature Computational Science},
      year={2024},
      month={Jan},
      day={01},
      volume={4},
      number={1},
      pages={43-56},
      issn={2662-8457},
      doi={10.1038/s43588-023-00573-5},
      url={https://doi.org/10.1038/s43588-023-00573-5}
}

@misc{life2vec_code,
  author = {Germans Savcisens},
  title = {Official code for the "Using Sequences of Life-events to Predict Human Lives" paper},
  note = {GitHub: SocialComplexityLab/life2vec},
  year = {2023},
  howpublished = {\url{https://doi.org/10.5281/zenodo.10118621}},
}