Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via Data Selection

Abstract: Machine learning models can fail on subgroups that are underrepresented during training. While techniques such as dataset balancing can improve performance on underperforming groups, they require access to training group annotations and can end up removing large portions of the dataset. In this paper, we introduce Data Debiasing with Datamodels (D3M), a debiasing approach which isolates and removes specific training examples that drive the model's failures on minority groups. Our approach enables us to efficiently train debiased classifiers while removing only a small number of examples, and does not require training group annotations or additional hyperparameter tuning.

• The paper introduces D3M, which leverages datamodels to compute influence scores and selectively remove training samples that degrade worst-group accuracy.
• It employs a method that calculates coefficients and group alignment scores to isolate problematic data, even inferring pseudo-group labels when explicit annotations are absent.
• Experimental results on biased datasets like CelebA and Waterbirds demonstrate significant improvements in subgroup robustness without substantial data reduction.

An Essay on "Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via Data Selection"

The paper "Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via Data Selection" presents an innovative technique for addressing subgroup robustness in machine learning models. The authors aim to improve model performance on underrepresented subgroups by selectively removing training examples that disproportionately contribute to model biases, all without requiring extensive group annotations or hyperparameter tuning.

Overview of the D3M Approach

The main contribution of the paper is the Data Debiasing with Datamodels (D3M) technique, which leverages the concept of datamodeling. Datamodeling approximates the model predictions as linear functions of the training data. Specifically, for a given validation example, a datamodel computes a set of coefficients that indicate the impact of each training example on the model's prediction for that validation example.

By aggregating these coefficients over subpopulations, D3M isolates training examples that most negatively impact worst-group accuracy (WGA). This subset of problematic examples is then removed from the training dataset, resulting in improved WGA.

Methodological Details

The D3M method proceeds through the following steps:

1. Coefficient Calculation: Using datamodeling, the method calculates a datamodel vector for each validation example. These vectors quantify the influence of each training example on the model's performance.
2. Group Alignment Scores: For each training sample, D3M calculates a group alignment score that measures its contribution to WGA. The authors propose a "soft maximum" function to weigh each group's loss relative to others, focusing removal on the samples that degrade WGA the most.
3. Selective Removal: The training examples with the most negative alignment scores are removed, and the model is retrained.

In scenarios where group labels are not available, the authors propose Auto-D3M. This method builds pseudo-group labels by projecting datamodel vectors onto the top principal component of a "datamodel matrix," which captures training sample dependencies. These pseudo-labeled groups are then used in the D3M procedure.

Experimental Evaluation

The authors validate D3M and Auto-D3M across several biased datasets, such as CelebA-Age, CelebA-Blond, Waterbirds, and MultiNLI. The results indicate that D3M consistently improves WGA compared to both ERM and existing specialized techniques, such as reweighting and data balancing. Notably, D3M achieves competitive worst-group accuracy without sacrificing the overall dataset size, removing significantly fewer training examples than naive balancing methods.

In scenarios without group labels, Auto-D3M demonstrates an effective way to discover and mitigate biases. The method identifies spurious correlations, such as color biases in ImageNet, and successfully enhances the model's robustness.

Implications and Future Directions

The research has several practical and theoretical implications. Practically, D3M offers a scalable and efficient solution for improving subgroup robustness in large, unannotated datasets. The method's ability to selectively debias without extensive annotation requirements makes it highly relevant for real-world applications where annotating large training datasets is infeasible.

Theoretically, D3M contributes to the understanding of how specific data points influence model fairness. By focusing on example-specific contributions to bias, the method provides insights into the nature of biases within datasets and how they propagate through model training.

Conclusion

Data Debiasing with Datamodels (D3M) stands out as an effective method for mitigating biases in machine learning models. Its data-centric approach aligns well with the growing need for practical debiasing techniques in AI. Future research could further explore the integration of D3M with model-based interventions and extend its application to other domains, ultimately leading to more robust and fair machine learning systems.