Spurious Correlations in Machine Learning

Machine learning systems are known to be sensitive to spurious correlations between biased features of the inputs (e.g., background, texture, and secondary objects) and the corresponding labels. These features and their correlations with the labels are known as "spurious" because they tend to change with shifts in real-world data distributions, which can negatively impact the model's generalization and robustness. In this survey, we provide a comprehensive review of this issue, along with a taxonomy of current state-of-the-art methods for addressing spurious correlations in machine learning models. Additionally, we summarize existing datasets, benchmarks, and metrics to aid future research. The paper concludes with a discussion of the recent advancements and future research challenges in this field, aiming to provide valuable insights for researchers in the related domains.

Here is a collection of research papers and the relevant valuable open-source resources for Spurious Correlations in Machine Learning. Watch this repository for the latest updates and feel free to raise pull requests if you find some interesting papers!

Our survey paper on arxiv: Spurious Correlations in Machine Learning: A Survey

@article{ye2024spurious,
  title={Spurious Correlations in Machine Learning: A Survey},
  author={Ye, Wenqian and Zheng, Guangtao and Cao, Xu and Ma, Yunsheng and Hu, Xia and Zhang, Aidong},
  journal={arXiv preprint arXiv:2402.12715},
  year={2024}
}

Table of Contents

• Datasets
• Articles
• Papers

Datasets

Dataset	#Class	#Sample	Brief Description	Domain
Colored MNIST	-	60,000	Synthetic Dataset	Vision
Corrupted CIFAR-10	-	60,000	Synthetic Dataset	Vision
CelebA	2x2	202,599	Face Attribute	Vision
Waterbirds	2x2	11,788	Bird Subset of CUB	Vision
ImageNet-9	9	5,495	Subset of ImageNet	Vision
ImageNet-A	200	7,500	Subset of ImageNet	Vision
ImageNet-C	75	-	Subset of ImageNet	Vision
BAR	6	2,595	Action	Vision
NICO	19	24,214	Animal and Vehicle	Vision
MetaShift	410	12,868	Natural Images	Vision
FMOW	63	1,047,691	Satellite Images	Vision
Spawrious	6x4	152,064	Dog	Vision
SpuCoMNIST	4	70,000	Synthetic Dataset	Vision
SpuCoAnimals	4	48,100	Natural Images	Vision
CivilComments	7	1,999,514	Public Comment	Text
MultiNLI	3x2	206,175	Language Inference	Text
FDCL18	7	80,000	Abusive Behavior	Text
IV-VQA	-	657,090	Causal VQA	VQA
CV-VQA	-	59,513	Causal VQA	VQA
ChestX-ray14	14	112,120	Chest X-Ray	Health
MIMIC-CXR	-	377,110	Chest X-Ray	Health
CheXpert	14	224,316	Chest X-Ray	Health
PadChest	19	160,868	Chest X-Ray	Health

Articles

• Spurious Correlations in Machine Learning
• Spurious Correlation: Definition, Examples & Detecting
• Spurious Correlation: Definition, How It Works, and Examples
• Why Correlation-Based Machine Learning Leads to Bad Predictions
• How Causal Inference Helps Machine Learning Models Avoid Spurious Correlations and Make Informed Decisions

Papers

• On Feature Learning in the Presence of Spurious Correlations
• An Investigation of Why Overparameterization Exacerbates Spurious Correlations

Multi-Modal

• Understanding the Robustness of Multi-modal Contrastive Learning to Distribution Shift