This repository contains the implementation of Data-IQ, a "Data-Centric AI" framework to characterize subgroups with heterogeneous outcomes in tabular data.

Data-IQ studies training dynamics and specifically studies the inherent data uncertainty (aleatoric uncertainty), to characterize data examples into the following subgroups: EASY, AMBIGUOUS and HARD.

Data-IQ can be used to characterize data using any machine learning model trained iteratively. We provide an interface for using Data-IQ, which can be used with Pytorch models, as well as, SKLearn style API models including XGBoost, LightGBM and CatBoost.

The utility of the subgroups extend to a variety of use-cases.

For more details, please read our NeurIPS 2022 paper: Data-IQ: Characterizing subgroups with heterogeneous outcomes in tabular data.

1. Clone the repository
2. Create a new virtual environment with Python 3.7. e.g:

    virtualenv dataiq_env

3. With the venv activated, run the following command from the repository directory:

• Minimum requirements to run Data-IQ on your own data

pip install data_iq

or from source,

pip install .

• Full requirements to run Data-IQ tests and tutorials

pip install .[testing]

4. Link the venv to the kernel:

python -m ipykernel install --user --name=dataiq_env

• Option 1: Install as a package (called data_iq) from PyPI using

pip install data_iq

or from source using

   python -m pip install -e .

• Option 2: Import from the data_iq folder

Two examples using a neural network (Pytorch-style) and XGBoost (Sklearn-style) is shown below.

The lines of Data-IQ code that need to be added to your training loops are highlighted with stars in the comments. e.g. *** comment ***

# Import Data-IQ
# Option 1 - if package is installed
from data_iq import DataIQ_Torch

# Option 2 - import from folder
from data_iq.dataiq_class import DataIQ_Torch

# Pytorch data loader here
train_loader = DataLoader(dataset=train_data, batch_size=128, shuffle=True)

# Define Pytorch device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Define Pytorch model
net = Example_NN(input_size=X_train.shape[1], nlabels=len(np.unique(y_train)))
net.to(device)

# *** Initialize Data-IQ [LINE 1] ***
dataiq = DataIQ_Torch(X=X_train, y=y_train, sparse_labels=True)



# Fit Pytorch model [Training loop]
for e in range(1, EPOCHS + 1):
    net.train()

    for X_batch, y_batch in train_loader:
        ### ADD TRAIN LOOP HERE ###


    # *** CALL DATA-IQ on EPOCH END [LINE 2] ***
    dataiq.on_epoch_end(net, device=device)



# *** Access metrics ***
aleatoric_uncertainty = dataiq_xgb.aleatoric
confidence = dataiq_xgb.confidence

# Import Data-IQ
# Option 1 - if package is installed
from data_iq import DataIQ_SKLearn

# Option 2 - import from folder
from data_iq.dataiq_class import DataIQ_SKLearn

# Arbitrary data loader - numpy arrays
X_train, X_test, y_train, y_test = load_data()

# *** Initialize Data-IQ [LINE 1] ***
dataiq_xgb = DataIQ_SKLearn(X=X_train, y=y_train)

# Fit XGBoost
clf = xgb.XGBClassifier(n_estimators=10)
clf.fit(X_train, y_train)

for i in range(1, nest):
    # *** Characterize with Data-IQ [LINE 2] ***
    dataiq_xgb.on_epoch_end(clf=clf, iteration=i)

# *** Access metrics ***
aleatoric_uncertainty = dataiq_xgb.aleatoric
confidence = dataiq_xgb.confidence

To get started with Data-IQ on your own data, we provide two tutorial notebooks to illustrate the usage of Data-IQ. Examples are provided both Pytorch style and SKLearn style (XGBoost, Catboost, LightGBM) models.

These notebooks can be found in the /tutorial folder. The Adult open-source dataset is given as an example dataset in the tutorials for ease of accessbility.

1. tutorial_torch_api.ipynb

• Example integration of Data-IQ with Pytorch models

2. tutorial_sklearn_api.ipynb

• Example integration of Data-IQ with SKLearn style models trained iteratively (e.g. XGBoost, LigthGBM, CatBoost)

The Data-IQ package provides a wide variety of metrics to characterize training dynamics.

The primary & suggested metrics are: Aleatoric uncertainty and Predictive confidence. However, other metrics are also included as part of the package.

The different metrics available to characterize training dynamics are as follows and can be accessed through the dataiq object:

• Aleatoric uncertainty via: dataiq.aleatoric
• Predictive confidence via: dataiq.confidence
• Variability via: dataiq.variability
• Entropy via: dataiq.entropy
• Mutual information via: dataiq.mi
• Correctness over training via: dataiq.correctness

We highlight different use-cases of Data-IQ from understanding learning dynamics, creating characteristic curves, feature acquisition etc as well as, different data modalities in notebooks which can be found in the /use_cases folder.

If you use this code, please cite the associated paper:

@inproceedings
{seedat2022dataiq,
title={Data-IQ: Characterizing subgroups with heterogeneous outcomes in tabular data},
author={Seedat, Nabeel and Crabbe, Jonathan and Bica, Ioana and van der Schaar, Mihaela},
booktitle={Advances in Neural Information Processing Systems},
year={2022}
}