RVAE for Mixed Type Features (Tabular Data)

Code for paper: "Robust Variational Autoencoders for Outlier Detection and Repair of Mixed-Type Data" AISTATS 2020. Check it here https://arxiv.org/abs/1907.06671. Please consider citing us if you use our code.

Instalation

• Please install ./setup.py in folder ./src in order to use core_models package.

• Use Pytorch 1.3.1 at least

Usage

Data folder

• Clean is found (or inserted) in data_simple:
  • e.g. ./data_simple/Wine/Wine.csv
  • the folder contains both clean data, and then after nosing, several noisy replicas given by python noising_process.py in separate folders.

Output folder

• Current scripts generate folder ./outputs_experiments_i/{dataset}/{noise_type}/{corruption_level}_run_j/{Model_Name}/
• Therein results for outlier detection metrics (cell and row), and repair of cells are presented. The latter only if algorithm provides this.

Simple Work Flow

Noising a dataset:

• Go to ./src/dataset_prep_simple/ to run noising of datasets in data folder:
  • Open noising_process.py
  • Edit script definitions: dataset; noise type; corruption level;
  • Run python noising_process.py

Running a model:

• Go to ./src/scripts/ to run a specific model (choose from scripts therein):
  • Make sure you pick correct hyper-parameters (see paper), and turn --cuda-on for GPU. For instance:
    • sh run_RVAE_CVI.sh , for our main algorithm.
    • sh run_VAE_l2.sh , for VAE-L2 baseline.
    • sh run_CondPred.sh , for NN-based Conditional Predictor (pseudo-likelihood).
    • sh run_baselines.sh , for assorted baselines in paper.

License

MIT