GAUGE：面向不完整多模态分类的粒度自适应反事实证据门控

文章背景与核心概要

在现实世界的机器学习应用中，多模态分类任务经常面临输入缺失（Incomplete Multimodal）的挑战。现有的解决方法（如数据插补和动态融合）通常在较粗的模态整体级别上进行操作，这导致它们难以在同一个恢复出的模态内部同时保留可靠成分并抑制误导性成分，从而最终影响了预测结果的可靠性。

为了突破这一局限，本文作者推出了 GAUGE，这是一个专为不完整多模态分类设计的轻量级反事实门控框架。GAUGE 能够将观察到和恢复出的输入统一编码为细粒度的证据单元，利用预测感知的泰勒证据得分高效评估反事实影响，并通过加法注意力调制实现单元级的证据调节，且无需修改底层骨干网络架构。大量的实验和理论分析证明了该方法的优越性与准确性。

GAUGE: Granularity-Adaptive Counterfactual Gating of Evidence for Incomplete Multimodal Classification

Authors: Yunping Shi, En Yu, Kairui Guo, Jie Lu
Published: August 6, 2026
Primary Subject: Machine Learning (cs.LG)
arXiv: 2608.05608

GAUGE: Granularity-Adaptive Counterfactual Gating of Evidence for Incomplete Multimodal Classification

Authors: Yunping Shi, En Yu, Kairui Guo, Jie Lu
Published: August 6, 2026
Primary Subject: Machine Learning (cs.LG)
arXiv: 2608.05608

Summary

现实世界的多模态分类任务经常遇到输入不完整的情况。虽然现有的方法（如插补和动态融合）试图解决缺失模态的问题，但它们通常在粗粒度的、模态整体的级别上运行。因此，它们很难在同一个恢复出的模态内保留可靠的成分同时抑制误导性的成分，这最终损害了预测的可靠性。

为了克服这一局限性，作者引入了 GAUGE，这是一个专为不完整多模态分类设计的轻量级反事实门控框架。

Summary

Real-world multimodal classification tasks frequently encounter incomplete inputs. While existing methods—such as imputation and dynamic fusion—attempt to resolve missing modalities, they typically operate at a coarse, modality-wide level. Consequently, they struggle to retain reliable components while suppressing misleading ones within the same recovered modality, which ultimately compromises prediction reliability.

To overcome this limitation, the authors introduce GAUGE, a lightweight counterfactual gating framework designed for incomplete multimodal classification.

GAUGE 的核心贡献：

细粒度证据单元： 使用冻结的插补器来处理缺失模态，并将观察到的输入和恢复出的输入统一编码为细粒度的证据单元。
反事实效应打分： 利用预测感知的泰勒证据得分（prediction-aware Taylor evidence scores），评估用参考表示替换每个证据单元的反事实效应，该得分可在单次前向-后向传递中高效计算。
加法注意力调制： 将计算出的得分映射为连续的门（gates），将其转化为单元级证据调节的加法注意力-Logit 偏置，而无需对底层骨干网络架构进行任何修改。
强有力的实证与理论验证： 在六个基准测试上的广泛实验表明，在各种不完整输入设置下，GAUGE 始终优于强大的基线模型。此外，泰勒余项理论分析验证了一阶近似相对于精确反事实效应的准确性。

Key Contributions of GAUGE:

Fine-Grained Evidence Units: Uses a frozen imputer to handle missing modalities and uniformly encodes both observed and recovered inputs as fine-grained evidence units.

Counterfactual Effect Scoring: Evaluates the counterfactual effect of replacing every evidence unit with a reference representation using prediction-aware Taylor evidence scores, efficiently computed in a single forward-backward pass.

Additive Attention-Modulation: Maps the computed scores to continuous gates, translating them into additive attention-logit biases for unit-wise evidence modulation without requiring any modifications to the underlying backbone architecture.

Strong Empirical & Theoretical Validation: Extensive experiments across six benchmarks demonstrate that GAUGE consistently outperforms strong baselines across diverse incomplete-input settings. Additionally, a Taylor remainder theoretical analysis validates the accuracy of the first-order approximation against the exact counterfactual effect.