DCAST: Diverse Class-Aware Self-Training Mitigates Selection Bias for Fairer Learning

TL;DR

Diverse Class-Aware Self-Training (DCAST), model-agnostic mitigation aware of class-specific bias, which promotes sample diversity to counter confirmation bias of conventional self-training while leveraging unlabeled samples for an improved representation of the underlying population is introduced.

Abstract

Fairness in machine learning seeks to mitigate model bias against individuals based on sensitive features such as sex or age, often caused by an uneven representation of the population in the training data due to selection bias. Notably, bias unascribed to sensitive features is challenging to identify and typically goes undiagnosed, despite its prominence in complex high-dimensional data from fields like computer vision and molecular biomedicine. Strategies to mitigate unidentified bias and evaluate mitigation methods are crucially needed, yet remain underexplored. We introduce: (i) Diverse Class-Aware Self-Training (DCAST), model-agnostic mitigation aware of class-specific bias, which promotes sample diversity to counter confirmation bias of conventional self-training while leveraging unlabeled samples for an improved representation of the underlying population; (ii) hierarchy bias, multivariate and class-aware bias induction without prior knowledge. Models learned with DCAST showed improved robustness to hierarchy and other biases across eleven datasets, against conventional self-training and six prominent domain adaptation techniques. Advantage was largest on multi-class classification, emphasizing DCAST as a promising strategy for fairer learning in different contexts.

Figures (6)

• Figure 1: Hierarchy bias approach for induction of selection bias. Given input data $\boldsymbol{X}$ with labels $\boldsymbol{Y}$, number of samples to select $k$, and bias ratio $b \in [0,1]$, hierarchy bias selects $k$ samples per class $c$: $k \times b$ from a specific group and $k \times (1-b)$ from the remaining samples. Each class-specific candidate group (for class $c$) is identified via agglomerative hierarchical clustering with Euclidean distances and Ward linkage of the $c$-labeled samples until a cluster of size $\geq k$ is obtained, from which $k \times b$ samples are drawn uniformly at random. The $k \times (1-b)$ samples are drawn uniformly at random from the remaining $c$-labeled samples.
• Figure 2: Bias induction impact on sample distances, latent space, and classifier performance.(a) Class-specific distributions of per sample average Euclidean distances to all other samples, for the biased selection (histograms) and for all samples in the labeled train set (histogram peaks denoted by lines ending in a "T" shape), using three bias induction techniques (hierarchy with $b=0.9$, joint, and Dirichlet) and random subsampling on three datasets (wine, mushroom, and fire). Kolmogorov-Smirnov (KS) effect sizes quantify the distribution shift between the biased selection vs. all samples. (b-d) Samples selected by hierarchy bias ($b=0.9$), highlighted on the respective latent UMAP space of the labeled train set for the wine, mushroom, and fire datasets (arbitrarily chosen run 11). (e) Accuracy of supervised RF, NN, and LR models on the test set after training on the original or biased labeled train set, over 30 distinct train runs. Box height delimits the interquartile range ($IQR = Q3-Q1$), with a line across the box denoting the median; whiskers indicate the largest and smallest values within $Q1-1.5 \times IQR$ and $Q3+1.5 \times IQR$, with points beyond the range as outliers.
• Figure 3: Diverse Class-Aware Self-Training (DCAST) framework. (Left) Input to DCAST. Labeled data $\boldsymbol{X_L}$ (with labels $\boldsymbol{Y_L}$) and unlabeled data $\boldsymbol{X_U}$, maximum number of iterations $m$, number of pseudo-labeled samples $s$ to select per iteration, confidence or prediction probability threshold $t \in [0,1]$, and integer diversity strength parameter $d \geq 1$. (Middle) Self-training module. At each iteration, a model trained with labeled samples is used to predict pseudo-labels for unlabeled samples, from which a subset is newly selected and added to the labeled set for the next iteration. (Right) Diversity module. Selects the subset of $s_c = s \times class\_ratio(c)$ confidently predicted and diverse pseudo-labeled samples per class $c$, as follows: (i) select the top $s_c \times d$ samples from the unlabeled set with confidence or prediction probability larger than $t$ (or $1.2/C$, whichever is largest); and (ii) reduce this $s_c \times d$ selection to a set of $s_c$ diverse samples by identifying $s_c$ clusters using hierarchical clustering (agglomerative single-linkage) and selecting the most confidently predicted sample from each cluster. Note that $class\_ratio$ can otherwise be fixed to be equal across classes. Distance between samples is based on either learned discriminative embeddings, relating samples with respect to prediction output, or alternatively an unsupervised embedding or the original feature space. When $d=1$, DCAST becomes CAST, without the diversity strategy.
• Figure 4: Bias mitigation by semi-supervised (D)CAST in the presence of hierarchy bias (ratio $b=0.9$). Accuracy of supervised and semi-supervised learning methods with (a) RF, (b) NN, and (c) LR models across 11 datasets. Results for 30 runs: each training on a different split of the train set into labeled and unlabeled sets, all evaluated on the same original test set. Models included (top to bottom): supervised RF/NN/LR models trained on the original (No Bias) or biased (Bias) labeled set; and semi-supervised RF/NN/LR models, using conventional self-training (ST) on the biased labeled train set plus the unlabeled test set, or (D)CAST on the biased labeled train set plus the unlabeled train set. Red asterisks (*) denote statistically significant changes in accuracy over 30 runs for each semi-supervised approach compared to supervised learning on the biased labeled set, using one-sided Wilcoxon signed-rank tests (larger asterisks indicate $p < 0.01$ and smaller asterisks $0.01 < p < 0.05$).
• Figure 5: Bias mitigation by (D)CAST or domain adaptation beyond semi-supervised learning under hierarchy bias ($b=0.9$). Accuracy of semi-supervised (D)CAST strategies against alternative bias mitigation techniques with 3 different types of ML models for 11 datasets over 30 runs. Per run, each model was trained using a different labeled train set with induced hierarchy bias. We included a supervised learning model as baseline per ML model type (RF, NN, LR), together with bias mitigation models incorporating additional unlabeled samples from either the unlabeled train set ((D)CAST) or the unlabeled test set (remaining methods). All models were evaluated on the same original test set. Bias mitigation methods per category: semi-supervised (CAST and DCAST-100); importance weighting (KMM, KDE); minimax estimation (RBA, TCPR); and subspace alignment (FLDA, SUBA). The (D)CAST and KMM methods were coupled with RF, NN, and LR models, whereas the remaining methods used LR models only. For clarity, horizontal lines group bias mitigation strategies by model type. The "x" symbol indicates model training was unsuccessful across all 30 runs.