AutoCure: Automated Tabular Data Curation for ML Pipelines

The repository comprises the source code of the AutoCure method which implements an adaptive ensemble-based error detection method followed by a data augmentation approach to automatically curate tabular data for predictive tasks.

Setup

clone with submodules

https://audio.digitalbusinessplatform.de/gitlab/kompaki/norepair4ml/augmentation-data-cleaning.git

Install requirements

python3 -m venv venv 
source venv/bin/activate
pip3 install --upgrade setuptools
pip3 install --upgrade pip
pip3 install -e .

Install error detection and repair methods

RAHA and BARAN

To install these methods, you can do so in two different ways:

Option 1: through pip3

pip3 install raha

Option 2: through the setup.py script which exists in the raha main directory

python3 setup.py install

Katara

For this method, we do not need to install packages, but we need to download the knowledge base: Download the knowledge base (link) and unzip the file. The files of the knowledge base should be placed in the following path.

cd detectors/katara/knowedge-base

FAHES

To install FAHES, navigate to the src directory and run make to compile the source code

cd FAHES/src/

make clean && make

HoloClean

To install HoloClean, read te installation part of its README file.

Usage

The executable python scripts are placed in the scripts folder.

Run the threshold experiment

This command runs an experiment to examine the impact of changing the voting threshold on the detection recall and precision. In this experiment, we use the traditional Min-K detection method to find errors in dirty data sets. The results are stored in the directory experiments/evaluation/data/test_threshold/{data set name}. Several data sets van be employed at the same time, while the option verbose can be used to print intermediate results and to show the execution progress.

 python3 scripts/test_threshold.py 
        --dataset_name adult housing 
        --verbose

Run the pipeline with the clean data set

This command trains a neural network on the clean versions of the data sets, i.e., the ground truth. The option tune_hyperparams activates the Optuna module to optimize the hyperparameters. While, the option epochs defines the number of epochs while training the neural network. To estimate the variance and average values, the option nb_iterations is used to define how many times the experiment will be repeated. The results are stored in the directory experiments/evaluation/data/modeling/{data set name}. For each run, a figure of the learning curves is stored in the directory experiments/evaluation/plots/learning_curve/{data set name}. It is important to highlight that this experiment can be executed, even in the lack of the ground truth version of the data set.

 python3 scripts/test_clean.py 
        --dataset_name adult housing 
        --nb_iterations 10
        --epochs 500
        --tune_hyperparams
        --verbose

Run AutoCue

This command runs a pipeline composed of the AutoCure method followed by a neural network with regression, binary classification, or multi-class classification, according to the data set. The option nb_generated_samples defines the amount of clean data to be augmented by AutoCure. The results are stored in the directory experiments/evaluation/data/modeling/{data set name}.

 python3 scripts/test_augclean.py 
        --dataset_name adult housing 
        --nb_iterations 10
        --nb_generated_samples 2000
        --epochs 500
        --tune_hyperparams
        --verbose

Run the baseline methods

This command runs an end-to-end pipeline consists of an error detection method, a repair method, and a neural network. The list of available error detection methods include: IF, SD, IQR, mvdetector, raha, ed2, dBoost, min_k, holoclean, fahes, katara, nadeef. While, the list of repair methods involve baran, cleanwithGroundTruth, standardImputer, mlImputer, activeClean. The results are stored in the directory experiments/evaluation/data/modeling/{data set name}.

 python3 scripts/test_e2e.py 
        --dataset_name adult
        --nb_iterations 10
        --detection_method raha ed2 katara 
        --repair_method mlImputer baran
        --epochs 500
        --tune_hyperparams
        --verbose

Run the augmentation experiment

This command runs the AutoCure experiment with gradually increasing the amount of augmented clean data. The results are stored in the directory experiments/evaluation/data/augmentation/{data set name}.

 python3 scripts/test_augmentation.py 
        --dataset_name adult housing 
        --nb_iterations 10
        --epochs 500
        --tune_hyperparams
        --verbose

Run the robustness experiment

This command runs the AutoCure method or the baseline methods while increasing the amount errors in the dirty data set. The option experiment_type is used to select between running AutoCure or the baseline methods. Specifically, it has two valid values, namely e2e and aug2clean. The results are stored in the directory experiments/evaluation/data/robustness/{data set name}.

 python3 scripts/test_error_rates.py 
        --dataset_name adult
        --nb_iterations 10
        --experiment_type e2e
        --detection_method raha ed2 katara 
        --repair_method mlImputer baran
        --nb_generated_samples 2000
        --epochs 500
        --tune_hyperparams
        --verbose

Adding a dataset

To add a new dataset, follow the steps below:

• Add the CSV files of the dirty and the ground truth in the directory experiments/data/{data set name}/. Rename the CSV file of the ground truth to clean.csv and the CSV file of the dirty version to dirty.csv.
• Create a YML file to store metadata about the newly-added data set. The YML file has the following fields which have to be defined. The field name defines the name of the data set, while the field ml_task is used to define the machine learning task associated with this data set. Possible ML tasks include binary_classification, multiclass_classification, and regression.

name: adult
ml_task: binary_classification
labels: income
fd_constraints:

• To run the rule-based baseline methods, the following cleaning signals have to be provided for the new dataset:
  • Functional dependency (FD rules) and patterns to run NADEEF
  • Denial constraints (DC rules) for HoloClean

Experiments

In this section, we list all experiments planned in this project.

• test the performance of autoCure in terms of
  • execution time
  • performance of downstream ML models
• compare AutoCure with a set of baselines
• test robustness against different error types
• test the amount of augmented data

Todos

• implement an ensemble-based error detection
  • implement a dataset class for loading, preparing, storing data, etc.
  • implement baseline to show how selecting data repair requires expertise
    • implement an ensemble detector which will be used for AutoCure and for the baselines
    • implement repair methods
• implement the adaptive data sampler
• implement the data augmentation method

Follow-up Ideas

• integrate a data valuation method
• integrate an automatic tool for the generation of functional dependency rules