# Node attribute imputation using variational inference (NAIVI)

Author: **Simon Fontaine** (simfont@umich.edu)


## Requirements

All requirements are contained in the `requirements.txt` file, or can be obtained using the `poetry` library.

Note: the latest version of `pypet` conflicts with newer versions of `numpy`. There are some basic types that were
removed from `numpy` that are referenced in `pypet`. Therefore, I am using a patched version of `pypet` that is
compatible with the latest version of `numpy`. This patched version, rather than the public version, is referenced in the `requirements.txt` file,
so you should not have any issues with the installation.

## Repository structure

### NAIVI

Contains the source code to run various node attribute imputation methods.

- `vmp`: the proposed method (joint latent space model estimated with variational message passing)
- `mle`: the joint latent space model estimated using either maximum likelihood estimation with projection (MLE) or maximum a posteriori estimation (MAP) using the prior
- `mcmc`: the joint latent space model estimated using the NUTS sampler from Pyro
- `smoothing`: local averaging
- `mice`: imputation within the data matrix only using either MICE with linear models or KNN
- `constant`: imputation using the mean
- `gcn`: imputation using a graph convolutional network

### Pypet experiments

This folder contains utilities to run experiments using the `pypet` library.

### Experiments

This folder contains the scripts to run the experiments using `pypet`. Figures are also included in the
respective folders; raw results are omitted to reduce memory requirements for the repository, but are available upon request.

### Datasets

Datasets used in the real data experiments.

### Examples

#### Using the `pypet` manager

The file `single_experiment.py` shows how to run a single experiment using the `pypet` manager.