Recovering individual emotional states from sparse ratings using collaborative filtering

This repository contains all the code, data, and analyses supporting the associated manuscript. You may also be interested in the associated toolbox Neighbors.

Analysis

1. Run preprocess_decisions.py and preprocess_moth.py
2. Run fit_all_models.py 3 times once for each dataset: iaps,moth, anddecisions
3. Run fit_mice.py to run MICE across all 3 datasets
4. Examine analysis results in Model_Comparisons.ipynb notebook

Data

All .csv files in this repo are handled by Git LFS which needs to be installed before cloning/pulling this repo in order to populate these files with data. You can do this buy downloading git lfs from the link above and then typing the following command in a terminal: git lfs install. Afterwards, any git clone or git pull operations should automatically grab the large files from Github's large file storage. At anytime you can see what files are being managed by LFS using: git lfs ls-files.

Library functions

Additional helper functions used in various scripts are notebooks are located in lib/lib.py, and can be imported for use, e.g. from lib import verify_results

Python Environment Setup

The environment.yml file in this repo can be used to bootstrap a conda environment for reproducibility:

conda env create --name collab_filter --file environment.yml

Original setup

Project dependencies were managed using conda and pip. They have been exported using:

conda env export > environment.yml.

This environment was originally created using:

conda create --name collab_filter -c ejolly -c defaults -c conda-forge python=3.8 "r::r-base" pymer4.

Additional packages were installed manually using conda install or pip install.

Parallelization with Ray

fit_all_models.py requires a working Ray installation

1. start a ray "head-node" instance on the multi-core machine you plan to use, e.g. after ssh-ing

   ray start --head -v

2. Now run fit_all_models.py following the documentation in the script

3. Examine live estimation time and log messages at Ray's web dashboard server localhost:8125 (default). If you ssh'd with port-forwarding enable you should be able to see Ray's dashboard through your local machine's browser as well

Development setup

Development work was performed in Visual Studio Code, which requires it's own configuration to work well with the conda environment. Once your environment is setup, you should appropriately adjust the relevant paths in .vscode/settings.json for your system.