Missingness Bias Calibration in Feature Attribution Explanations

TL;DR

This paper introduces MCal, a lightweight post-hoc correction method that effectively reduces missingness bias in feature attribution explanations without retraining models, improving reliability across various domains.

Contribution

The paper challenges the view of missingness bias as a deep flaw and proposes MCal, a simple linear correction method that outperforms complex solutions.

Findings

MCal significantly reduces missingness bias in feature importance scores.

MCal is competitive with or better than heavyweight approaches.

Effective across vision, language, and tabular medical benchmarks.

Abstract

Popular explanation methods often produce unreliable feature importance scores due to missingness bias, a systematic distortion that arises when models are probed with ablated, out-of-distribution inputs. Existing solutions treat this as a deep representational flaw that requires expensive retraining or architectural modifications. In this work, we challenge this assumption and show that missingness bias can be effectively treated as a superficial artifact of the model's output space. We introduce MCal, a lightweight post-hoc method that corrects this bias by fine-tuning a simple linear head on the outputs of a frozen base model. Surprisingly, we find this simple correction consistently reduces missingness bias and is competitive with, or even outperforms, prior heavyweight approaches across diverse medical benchmarks spanning vision, language, and tabular domains.