Generalizable Data Cleaning of Tabular Data in Latent Space

Citation

Please use the following BibTeX code to cite our paper, which is accepted for presentation at VLDB 2024:

@inproceedings{lopster2024,  
title       = {Generalizable Data Cleaning of Tabular Data in Latent Space},  
author      = {Eduardo dos Reis, Mohamed Abdelaal and Carsten Binnig},  
journal     = {Proceedings of the VLDB Endowment},  
volume      = {17},  
number      = {13},  
pages       = {4786--4798},  
year        = {2024},  
publisher   = {VLDB Endowment}  
}

Data cleaning on your own dataset

We provided the training and cleaning in a single script lopster.py.

Step 1. The requirements are two csv files inside a folder < my dataset folder >: a clean.csv containing a - mostly - clean sample of the data, and dirty01.csv, the target dataset to be cleaned. Keep this folder in the project root folder for ease-of-use.

• < my dataset folder >
  • clean.csv
  • dirty01.csv

Step 2 Open the dataset_configuration.json file and replace the "empty" json entry with your dataset information, while also replacing the name of the entry for < my dataset folder >.

Next, the script will yield a lopster.csv file containing the full cleaned version of the dataset under < my dataset folder > . If it is the first time running the script for the current dataset, it will first train a Lopster model using tensorflow. Example usage:

python3 lopster.py --dataset < my dataset folder >

You can also change hyperparameters for training (K, batch size, latent space dimensionality, training epochs):

python3 lopster.py --dataset < my dataset folder > --path < path/to/model/folder/ > --K 12 --latent 120 --epochs 100 --batch_size 256

Setup for reproducibility

We provided the REIN benchmark data used for the paper as is, in rein_data.zip. Once the data is decompressed to the DATASETS_REIN/ folder, one must create a python 3 environment and install the dependencies listed on requirements.txt. Next, a new Lopster model can be trained on any dataset as long as its configuration is mapped to the datasets.py script.

Reproducing results and Training

To reproduce the paper results on the REIN benchmark, a script was provided (rein_benchmark.py). For example:

python3 rein_benchmark.py --dataset adult  --K 12 --latent 120 --epochs 100 --batch_size 256 --eval_tuples 30000

Ablation Studies

All our ablation studies are available on ablation_studies.py, and it is already configured to replicate only the published ones.

python3 ablation_studies.py --dataset soccer_PLAYER  --K 12 --latent 120 --epochs 40 --batch_size 256 --training_tuples 30000

Plots

All the evaluation files are saved inside the evaluation/ folder as .CSV and can be easily plotted. We provide a script to replicate our plots in plotter.py, the usage follows:

python3 plotter.py --experiment tuplewise  --dataset adult --y_title  --legend

Datasets

The largest dataset used for the paper experiments is available for download at Large Soccer Dataset. All the others are available in the DATASETS_REIN folder.