Source-linked AI summary
Simple data balancing achieves competitive worst-group-accuracy
Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, David Lopez-Paz
TL;DR
The paper asks how to train classifiers that perform well across groups when common benchmarks contain substantial class and group imbalances. It compares class and group balancing through subsampling or reweighting with existing worst-group-accuracy methods. Simple balancing reaches state-of-the-art performance while training faster and requiring no additional hyper-parameters, and group information matters most for validation-based model selection.
Problem
Worst-group-accuracy seeks classifiers that remain effective across groups under shifted group proportions, while common benchmarks exhibit substantial class and group imbalances.
Method
The paper studies four benchmarks and compares state-of-the-art methods with simple class and group balancing through data subsampling or reweighting.
Results
Simple data balancing baselines achieve state-of-the-art performance across four popular worst-group-accuracy benchmarks while training faster and requiring no additional hyper-parameters.
Takeaways & Limitations
Group balancing performs best, but class balancing remains competitive without attribute information, while attribute access is most critical for model selection.
Takeaways & Limitations
Simple balancing may work mainly when spurious correlations dominate the data; more nuanced cases may benefit from complex methods such as gDRO.
Abstract
from arXiv · showhide
We study the problem of learning classifiers that perform well across (known or unknown) groups of data. After observing that common worst-group-accuracy datasets suffer from substantial imbalances, we set out to compare state-of-the-art methods to simple balancing of classes and groups by either subsampling or reweighting data. Our results show that these data balancing baselines achieve state-of-the-art-accuracy, while being faster to train and requiring no additional hyper-parameters. In addition, we highlight that access to group information is most critical for model selection purposes, and not so much during training. All in all, our findings beg closer examination of benchmarks and methods for research in worst-group-accuracy optimization.
1. Introduction
Worst-group-accuracy addresses distribution shifts in group proportions, where performance must remain strong across class–attribute groups rather than only on average. The paper finds that simple class and group balancing can match strong methods while being faster and simpler, with group information especially valuable for validation-based model selection.
- Motivation: Worst-group-accuracy seeks classifiers that maximize test performance on the worst group when training and testing group proportions differ.Groups are defined by class–attribute combinations, with attributes potentially labeled or unlabeled.
- Motivation: Worst-group-accuracy reduces reliance on spurious correlations and relates to building fair machine learning classifiers.Spurious correlations discriminate classes only within specific groups, which can undermine generalization.
- Approach: Four common benchmarks exhibit class and group imbalance, motivating experiments with subsampling and reweighting to balance classes and groups.The paper studies CelebA, Waterbirds, MultiNLI, and CivilComments.
- Findings: Simple data balancing baselines achieve competitive performance across four benchmarks while training faster and requiring no additional hyper-parameters.The experiments compare balancing by subsampling or reweighting with state-of-the-art worst-group-accuracy methods.
- Findings: Balancing groups gives the best results, while simple class balancing remains powerful when attribute information is unavailable.This distinguishes the value of group information during training from settings where only class labels are accessible.
- Findings: Attribute information is most critical for model selection in validation, rather than during training.The paper recommends trying subsampling first because it is faster, less sensitive to regularization, and more stable during long training sessions.
- Implications: The findings motivate closer examination of benchmarks and methods for future worst-group-accuracy research.The authors also question whether collecting more data should be the default strategy for worst-group-accuracy optimization.
2. Popular worst-group-accuracy benchmarks
The paper examines four worst-group-accuracy benchmarks built around examples with inputs, class labels, and attribute labels. Across these datasets, class and group proportions are substantially imbalanced and often encode spurious class–attribute correlations.
- Benchmark structure: Each benchmark example is represented as a triplet containing an input, a class label, and an attribute label.The four datasets follow this common structure for defining groups.
- Datasets: CelebA classifies whether aligned celebrity-face images show blond hair, with sex as the attribute.Each face image carries multiple traits, while the benchmark uses blond-hair classification and male/female attributes.
- Datasets: Waterbirds classifies birds as waterbirds or landbirds, with natural versus non-natural habitat as the attribute.The images depict birds composited onto different backgrounds.
- Datasets: MultiNLI classifies sentence-pair relationships as contradiction, entailment, or neither, using negation words in the second sentence as the attribute.Negation words make the contradiction label more likely in this dataset.
- Datasets: CivilComments classifies online comments as toxic or non-toxic using demographic and identity-related attributes.The coarse benchmark attribute indicates whether any of eight listed identity attributes appears in a comment.
- Imbalance and correlations: Three of four datasets have large class imbalance, all have large group imbalance, and class probabilities shift substantially when conditioned on attributes.Examples include inverted class probabilities across Waterbirds attributes and strong negation–contradiction association in MultiNLI.
- Imbalance and correlations: These imbalances motivate balancing classes and groups to reduce spurious correlations before comparing simple baselines with robust learners.The paper notes that dominant groups can encourage algorithms to rely on spurious correlations and memorization.
3. Popular worst-group-accuracy methods
The paper reviews ERM, JTT, and gDRO as prominent approaches to worst-group-accuracy optimization, differing in their use of attribute information and loss weighting. It also situates these methods within a broader literature of robust learners.
- ERM minimizes empirical risk without using attribute labels.
- JTT first trains an ERM model, then upweights examples misclassified by that model before training a final ERM model.JTT does not use attribute labels.
- gDRO minimizes the maximum loss across groups using attribute labels.It dynamically weights each group according to its current error.
- The reviewed literature also includes robust learners that do not use attribute information, such as Learning from Failure, Spectral Decoupling, and GEORGE.
4. Simple data balancing baselines
The paper evaluates class- and group-balancing baselines through subsampling or reweighting, motivated by imbalances in common benchmarks. In a toy setting, balancing groups can reduce reliance on spurious features, while model selection remains important for reweighting.
- Four baselines balance classes or groups by subsampling or reweighting example-sampling probabilities.
- SUBY subsamples large classes to the size of the smallest class without using attribute labels.
- SUBG subsamples large groups to the size of the smallest group and uses attribute labels.
- RWY and RWG make mini-batches class-balanced or group-balanced in expectation, respectively.RWY does not use attribute labels, whereas RWG does.
- In the toy example, class labels depend on core and spurious attributes whose correlations differ in stability across training and test data.The core correlation remains invariant, while the spurious correlation varies.
- ERM mainly relies on the faster-learned spurious feature, whereas SUBG decorrelates it from labels and achieves good test worst-group-accuracy.RWG also solves the toy example, but test worst-group-accuracy degrades as training iterations increase, making model selection important.
- 10^-11 is the toy example's misclassification probability when using only the core feature.The passage states that the problem is separable using just x_core with high probability.
- The figure depicts predicted class probabilities over the spurious-feature/core-feature plane, with red and blue shades indicating the two classes.Noise features are not depicted; the heatmap uses the nearest training sample's noise vector for visualization.
5. Experiments
Across four benchmarks, the experiments compare simple class/group balancing with state-of-the-art methods and analyze their accuracy, speed, hyper-parameter sensitivity, and training dynamics. Balancing is generally competitive, while subsampling is more stable during long training and MultiNLI is a notable exception.
- Experimental setup: The study evaluates ERM, JTT, gDRO, SUBY, SUBG, RWY, and RWG on Waterbirds, CelebA, MultiNLI, and CivilComments.Image datasets use pretrained ResNet50 models, while NLP datasets use pretrained BERT models.
- Results: 1.7 points less than gDRO is SUBG’s average score, while RWY scores 2.1 points more than JTT on average.SUBY performs below RWY, whereas SUBG slightly outperforms RWG.
- Analysis of exceptions: gDRO surpasses the second-best method by 8.4 points on MultiNLI, where the spurious attribute is relatively mild rather than dominant.The authors attribute this exception to gDRO’s softer group-loss maximization, which can handle a mild spurious correlation more flexibly than strong balancing.
- Results: Balancing baselines match state-of-the-art performance within error bars while using the minimum four tuned hyper-parameters.The four parameters are learning rate, weight decay, batch size, and early-stopping epoch.
- Analysis of exceptions: SUBY consistently underperforms on these datasets, possibly because rare worst-group examples may not be sampled at all.The authors therefore do not recommend SUBY in practice.
- Hyper-parameter sensitivity: Group-aware methods are less sensitive to hyper-parameter choice, while learning rate is the most influential hyper-parameter overall.Their top-five worst-group-accuracy range is smaller, so fewer than 50 tuning runs may still identify good models.
- Training dynamics: RWY, RWG, and gDRO peak early and then degrade, whereas SUBG remains stable during long training sessions.Removing regularization also degrades RWG, RWY, and gDRO, but SUBG maintains its performance without regularization.
- Reweighting versus subsampling: Early-stopped reweighting can effectively subsample majority examples because minority examples are repeatedly sampled with replacement.For the best RWG model on CelebA, three epochs expose only 44% of majority examples on average.
6. Conclusion
Simple data balancing achieves state-of-the-art performance on four worst-group-accuracy benchmarks, with group balancing strongest and class balancing competitive without attribute information. The study also emphasizes validation-time group labels, favors subsampling over reweighting in several practical respects, and identifies open questions about benchmarks and further improvements.
- Simple data balancing baselines achieve state-of-the-art performance on four popular worst-group-accuracy benchmarks.
- Balancing groups gives the best worst-group-accuracy, while balancing class labels remains competitive without attribute information.
- Attribute information is critical in validation for model selection based on worst-group-accuracy, while hyper-parameter tuning under weak supervision remains open.
- The authors advocate trying data subsampling before reweighting because it trains faster, relies less on regularization, and remains more stable during long training sessions.
- The findings raise whether current worst-group-accuracy benchmarks express a real problem and whether simple balancing can be outperformed on them.
Appendix A. Supplementary material
The appendix presents supplementary visualizations and tables for comparing methods, hyperparameters, and model behavior. Figure 3 uses a cherry-picked sentence to illustrate how gDRO, RWG, and SUBG differ in word-level importance and classification.
- Table 5 lists the best hyperparameters for each dataset and method alongside the corresponding worst test accuracy.
- Table 5 selects top runs using worst validation accuracy and uses color to distinguish smaller from higher values.
- Figure 3 compares gDRO, RWG, and SUBG on a cherry-picked sentence.
- Green highlights show the norm of the gradient of the contradiction probability with respect to each input embedding, serving as a proxy for word importance.
- In the illustration, gDRO identifies “remembering” and correctly classifies the relation as contradiction, whereas the other methods fail.