TL;DR
FairSplit is a novel in-processing method for fair classification that learns to decompose a model's embedding space into two distinct parts: a "Fair Space" used for predictions and a "Sensitive Space" that isolates biased information. By giving sensitive attributes a dedicated "destination" outside the classifier's view, FairSplit achieves state-of-the-art fairness-accuracy trade-offs and, unlike many competitors, naturally handles multiclass tasks and non-binary sensitive attributes.
What Problem Does This Paper Address?
Machine learning models often inadvertently learn to rely on sensitive attributes (like race or gender) to make predictions, perpetuating historical biases. The core challenge in "fair classification" is reducing this reliance without destroying the model's predictive accuracy.
While many methods exist, the authors identify three major gaps:
- Binary Bias: Most existing "in-processing" methods (which add fairness constraints during training) are designed strictly for binary classification or binary sensitive attributes. Real-world data is rarely that simple.
- Lack of Theory: Many methods show empirical success but lack theoretical guarantees that their approach actually bounds fairness violations.
- Prediction-Level Focus: Most methods regularize the final output. FairSplit argues that acting directly on the representation (the embedding space) provides more robust guarantees.
Methods
FairSplit introduces a learnable binary mask $b$ that acts on the embedding vector $z$ produced by an encoder. This mask splits the dimensions of $z$ into two subspaces:
- Fair Space ($z_f$): Dimensions where $b_j = 1$. This is the only part of the embedding the classifier sees.
- Sensitive Space ($z_s$): Dimensions where $b_j = 0$. This space is explicitly designed to "absorb" sensitive information.
The training process is guided by a three-part objective:
- Structural Compression: This forces dimensions with low within-class variance into the Fair Space and dimensions with low within-group (sensitive attribute) variance into the Sensitive Space.
- HSIC Regularization: The authors use the Hilbert-Schmidt Independence Criterion (HSIC) to statistically decouple the Fair Space from sensitive attributes while simultaneously maximizing the dependence of the Sensitive Space on those attributes.
- Utility: A standard cross-entropy loss ensures the Fair Space remains highly predictive for the target task.
A key technical highlight is Proposition 3.1, which provides a closed-form solution for updating the mask, allowing the dimensionality of the Fair Space to be learned dynamically rather than fixed as a hyperparameter.
Datasets
The authors evaluated FairSplit across diverse modalities and complexities:
- Tabular: Adult (income prediction) and COMPAS (recidivism risk). COMPAS was used specifically to test intersectional fairness (gender × age).
- Images: CelebA-A (predicting attractiveness vs. gender) and UTKFace-AR.
- Multiclass/Non-Binary: UTKFace-AR was the primary benchmark for complex settings, featuring a 3-class age target and 3 racial groups as sensitive attributes.
- Foundation Models: Experiments also utilized FairFace and CelebA at higher resolutions to test transferability with frozen encoders like CLIP and DINOv2.
Key Insights and Ideas
- The "Destination" Principle: Most fairness methods try to suppress sensitive information. FairSplit instead gives it a "destination" ($z_s$). By providing a place for the model to store biased features, it becomes easier to keep the Fair Space ($z_f$) clean.
- Dimensionality as a Fairness Lever: The authors prove (Theorem 3.2) that the demographic parity gap is upper-bounded by the dimensionality of the Fair Space. This means that learning a smaller Fair Space directly tightens the theoretical fairness guarantee.
- Representation vs. Prediction: Because FairSplit cleans the representation, the resulting "Fair Space" can be used for downstream tasks the model wasn't even trained for. The authors demonstrated this by training a Fair Space for one attribute (e.g., "high cheekbones") and showing it remained fair when a new head was trained to predict a different attribute (e.g., "age").
Key Results
- Multiclass Superiority: On the UTKFace-AR dataset (3 classes, 3 sensitive groups), FairSplit achieved a Demographic Parity gap ($\Delta DP$) of 12.6, significantly lower than the baseline ERM (27.9), while maintaining competitive F1 scores.
- Intersectional Fairness: On COMPAS, FairSplit achieved the lowest Area Between CDF Curves (ABCC) of 1.2, compared to 6.7 for the next best method, proving its effectiveness in complex, imbalanced group settings.
- Efficiency: FairSplit adds only about 17-22% computational overhead compared to standard training, which is on par with or faster than many adversarial debiasing methods.
Limitations and Future Work
- Linear Head Assumption: The theoretical proof for the demographic parity bound currently assumes a linear classification head. While empirical results show FairSplit still works well with non-linear heads, the theoretical guarantee doesn't yet extend to them.
- Sensitive Labels Required: Like most in-processing methods, FairSplit requires sensitive attribute labels during training. The authors suggest future work could explore semi-supervised versions for when these labels are scarce.
- Privacy Concerns: While the Fair Space is "clean," the Sensitive Space explicitly retains sensitive information. In privacy-sensitive applications, the model parameters themselves would need to be secured to prevent leakage of the protected attributes.
Summary updated