DoubleML/multimodal_confounding

Dataset Card

Semi-synthetic dataset with multimodal confounding. The dataset is generated according to the description in DoubleMLDeep: Estimation of Causal Effects with Multimodal Data.

Dataset Details

Dataset Description & Usage

The dataset is a semi-synthetic dataset as a benchmark for treatment effect estimation with multimodal confounding. The outcome variable Y is generated according to a partially linear model $$Y={\theta }_0{D}_1+g_1(X)+\epsilon Y\; =\; \backslash theta\_0\; D\_1\; +\; g\_1(X)\; +\; \backslash varepsilon$$ with an constant treatment effect of $${\theta }_0=\mathrm{0.5.}\backslash theta\_0=0.5.$$ The target variables sentiment, label and price are used to generate credible confounding by affecting both Y and D_1. This confounding is generated to be negative, such that estimates of the treatment effect should generally be smaller than 0.5.

For a more detailed description on the data generating process, see DoubleMLDeep: Estimation of Causal Effects with Multimodal Data.

The dataset includes the corresponding target variables sentiment, label, price and oracle values such as cond_exp_y, l1, m1, g1. These values are included for convenience for e.g. benchmarking against optimal estimates, but should not be used in the model. Further, several tabular features are highly correlated, such that it may be helpful to drop features such as x, y, z.

An example looks as follows:

{'cond_exp_y': 2.367230022801451,
 'm1': -2.7978920933712907,
 'g1': 4.015536418538365,
 'l1': 2.61659037185272,
 'Y': 3.091541535115522,
 'D_1': -3.2966127914738275,
 'carat': 0.5247285289349821,
 'depth': 58.7,
 'table': 59.0,
 'price': 9.7161333532141,
 'x': 7.87,
 'y': 7.78,
 'z': 4.59,
 'review': "I really liked this Summerslam due to the look of the arena, the curtains and just the look overall was interesting to me for some reason. Anyways, this could have been one of the best Summerslam's ever if the WWF didn't have Lex Luger in the main event against Yokozuna, now for it's time it was ok to have a huge fat man vs a strong man but I'm glad times have changed. It was a terrible main event just like every match Luger is in is terrible. Other matches on the card were Razor Ramon vs Ted Dibiase, Steiner Brothers vs Heavenly Bodies, Shawn Michaels vs Curt Hening, this was the event where Shawn named his big monster of a body guard Diesel, IRS vs 1-2-3 Kid, Bret Hart first takes on Doink then takes on Jerry Lawler and stuff with the Harts and Lawler was always very interesting, then Ludvig Borga destroyed Marty Jannetty, Undertaker took on Giant Gonzalez in another terrible match, The Smoking Gunns and Tatanka took on Bam Bam Bigelow and the Headshrinkers, and Yokozuna defended the world title against Lex Luger this match was boring and it has a terrible ending. However it deserves 8/10",
 'sentiment': 'positive',
 'label': 6,
 'cut_Good': False,
 'cut_Ideal': False,
 'cut_Premium': True,
 'cut_Very Good': False,
 'color_E': False,
 'color_F': True,
 'color_G': False,
 'color_H': False,
 'color_I': False,
 'color_J': False,
 'clarity_IF': False,
 'clarity_SI1': False,
 'clarity_SI2': False,
 'clarity_VS1': False,
 'clarity_VS2': True,
 'clarity_VVS1': False,
 'clarity_VVS2': False,
 'image': <PIL.PngImagePlugin.PngImageFile image mode=RGB size=32x32>}

Dataset Sources

The dataset is based on the three commonly used datasets:

• Diamonds dataset
• IMDB dataset
• CIFAR-10 dataset

The versions to create this dataset can be found on Kaggle:

• Diamonds dataset (Kaggle)
• IMDB dataset (Kaggle)
• CIFAR-10 dataset (Kaggle)

The original citations can be found below.

Dataset Preprocessing

All datasets are subsampled to be of equal size (50,000). The CIFAR-10 data is based on the trainings dataset, whereas the IMDB data contains train and test data to obtain 50,000 observations. The labels of the CIFAR-10 data are set to integer values 0 to 9. The Diamonds dataset is cleaned (values with x, y, z equal to 0 are removed) and outliers are dropped (such that 45<depth<75, 40<table<80, x<30, y<30 and 2<z<30). The remaining 53,907 observations are downsampled to the same size of 50,000 observations. Further price and carat are transformed with the natural logarithm and cut, color and clarity are dummy coded (with baselines Fair, D and I1).

Uses

The dataset should as a benchmark to compare different causal inference methods for observational data under multimodal confounding.

Dataset Structure

Data Instances

Data Fields

The data fields can be devided into several categories:

• Outcome and Treatments

  • Y (float64): Outcome of interest
  • D_1 (float64): Treatment value

• Text Features

  • review (string): IMDB review text
  • sentiment (string): Corresponding sentiment, either positive or negative

• Image Features

  • image (image): Image
  • label (int64): Corresponding label from 0 to 9

• Tabular Features

  • price (float64): Logarithm of the price in US dollars
  • carat (float64): Logarithm of the weight of the diamond
  • x (float64): Length in mm
  • y (float64): Width in mm
  • z (float64): Depth in mm
  • depth (float64): Total depth percentage
  • table (float64): Width of top of diamond relative to widest point
  • Cut: Quality of the cut (Fair, Good, Very Good, Premium, Ideal) (dummy coded with Fair as baseline)
    • cut_Good (bool)
    • cut_Very Good (bool)
    • cut_Premium (bool)
    • cut_Ideal (bool)
  • Color: Diamond color, from J(worst) to D(best) (dummy coded with D as baseline)
    • color_E (bool)
    • color_F (bool)
    • color_G (bool)
    • color_H (bool)
    • color_I (bool)
    • color_J (bool)
  • Clarity: Measurement of diamond clarity (I1 (worst), SI2, SI1, VS2, VS1, VVS2, VVS1, IF (best)) (dummy coded with I1 as baseline)
    • clarity_SI2 (bool)
    • clarity_SI1 (bool)
    • clarity_VS2 (bool)
    • clarity_VS1 (bool)
    • clarity_VVS2 (bool)
    • clarity_VVS1 (bool)
    • clarity_IF (bool)

• Oracle Features

  • cond_exp_y (float64): Expected value of Y conditional on D_1, sentiment, label and price
  • l1 (float64): Expected value of Y conditional on sentiment, label and price
  • m1 (float64): Expected value of D_1 conditional on sentiment, label and price
  • g1 (float64): Additive component of Y based on sentiment, label and price (see Dataset Description)

Limitations

As the confounding is generated via original labels, completely removing the confounding might not be possible.

Citation Information

Dataset Citation

If you use the dataset please cite this article:

@article{klaassen2024doublemldeep,
  title={DoubleMLDeep: Estimation of Causal Effects with Multimodal Data},
  author={Klaassen, Sven and Teichert-Kluge, Jan and Bach, Philipp and Chernozhukov, Victor and Spindler, Martin and Vijaykumar, Suhas},
  journal={arXiv preprint arXiv:2402.01785},
  year={2024}
}

Dataset Sources

The three original datasets can be cited via

Diamonds dataset:

@Book{ggplot2_book,
  author = {Hadley Wickham},
  title = {ggplot2: Elegant Graphics for Data Analysis},
  publisher = {Springer-Verlag New York},
  year = {2016},
  isbn = {978-3-319-24277-4},
  url = {https://ggplot2.tidyverse.org},
}

IMDB dataset:

@InProceedings{maas-EtAl:2011:ACL-HLT2011,
  author    = {Maas, Andrew L.  and  Daly, Raymond E.  and  Pham, Peter T.  and  Huang, Dan  and  Ng, Andrew Y.  and  Potts, Christopher},
  title     = {Learning Word Vectors for Sentiment Analysis},
  booktitle = {Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies},
  month     = {June},
  year      = {2011},
  address   = {Portland, Oregon, USA},
  publisher = {Association for Computational Linguistics},
  pages     = {142--150},
  url       = {http://www.aclweb.org/anthology/P11-1015}
}

CIFAR-10 dataset:

@TECHREPORT{Krizhevsky09learningmultiple,
    author = {Alex Krizhevsky},
    title = {Learning multiple layers of features from tiny images},
    institution = {},
    year = {2009}
}

Dataset Card Authors

Sven Klaassen