BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Problem

This work addresses the evaluation of trained predictors and the behavior of frozen backbones when learning new heads. It highlights the need for effective monitoring of spurious feature reliance, particularly in the context of class-attribute correlations. The paper is a preprint and has not undergone peer review.

Method

The authors propose a toolkit named BiasFlow, which includes functionalities for monitoring class-attribute centroid alignment (IBMI), within-class centroid separation (W-IBMI), and feature-projection sensitivity. The core of the method is BiasFlow Regularization (BFR), which is a supervised, composable class-conditional centroid-alignment penalty designed to mitigate reliance on spurious features. The metrics used include:

• IBMI: This metric is confounded by class-attribute correlation and does not measure causal feature reliance.
• W-IBMI: This metric is scale-dependent and verifies the quantity optimized by BFR.

Results

The paper reports several key results:

• A mean WGA Improvement of up to +26.0 percentage points on the UrbanCars dataset.
• An improvement in WGA when combining BFR with GroupDRO, increasing from 40.7% to 64.1%.
• A decrease in Male Probe Accuracy from 92.5% to 72.2%.
• An improvement in Watermark-Shift Accuracy of +23.0 percentage points under matched training conditions. The available text does not report quantitative results for baselines against which these improvements are measured.

Limitations

The authors note that W-IBMI does not independently establish attribute removal, which may limit its effectiveness in certain contexts. Additionally, they report mixed results in cross-task evaluations, indicating that the method may not generalize well across different tasks or datasets.

Why it matters

The implications of this work are significant for downstream applications in machine learning, particularly in scenarios where spurious feature reliance can lead to biased predictions. By providing a toolkit for monitoring and regularizing backbone models, BiasFlow aims to enhance the robustness and interpretability of machine learning systems, paving the way for more reliable deployment in real-world applications.