跳转至

文章背景与核心概要

在当前的机器学习领域,基于重构的异常检测器通常表现出强大的性能和准确性,但它们往往作为“黑盒”运行——在标记异常样本时,无法解释究竟是哪些特定的特征范围导致了异常。这种黑盒特性在金融、医疗和工业制造等对可解释性要求极高的场景中构成了巨大的应用障碍。

为了弥合高性能与人类可解释性之间的鸿沟,本文引入了 DIFFINT,这是一种创新的自编码器,其特征在于一个结构化为一组软轴对齐区间成员关系的潜空间瓶颈。该模型直接从原始数值数据中进行端到端学习,无需离散化或二值化处理,从而在保持深度学习强大表征能力的同时,维持了一个透明且可审计的内部结构。

本研究在 48 个 ADBench 基准测试中对 22 个基线模型进行了全面评估。结果表明,DIFFINT 不仅在 ROC-AUC 和 AUPR 指标上取得了整体最佳的平均排名,而且还是统计性能领先的七个方法集群中唯一的、具备可解释性的检测器,展现出了卓越的学术价值和应用潜力。


Differentiable Interval Bottlenecks for Interpretable Anomaly Detection in Numerical Data

Authors: Lamine Diop, Marc Plantevit
Submitted On: September 3, 2026
Accepted At: ICDM 2026
Primary Subject: Machine Learning (cs.LG)
arXiv ID: 2609.03878 [cs.LG]
DOI: 10.48550/arXiv.2609.03878

Authors: Lamine Diop, Marc Plantevit
Submitted On: September 3, 2026
Accepted At: ICDM 2026
Primary Subject: Machine Learning (cs.LG)
arXiv ID: 2609.03878 [cs.LG]
DOI: 10.48550/arXiv.2609.03878


📋 Summary

Reconstruction-based anomaly detectors are typically powerful and accurate, but they operate as "black boxes"—flagging anomalous samples without explaining which specific feature ranges caused the anomaly.

To bridge the gap between high performance and human interpretability, this paper introduces DIFFINT, an innovative autoencoder featuring a latent bottleneck structured as a set of soft, axis-aligned interval memberships. Learned end-to-end directly from raw numerical data without discretization or binarization, DIFFINT retains the representational power of deep learning while maintaining a transparent, auditable internal structure.

基于重构的异常检测器通常功能强大且准确,但它们的运行方式如同“黑盒”——在标记异常样本时,无法解释是哪些具体的特征范围导致了异常。

为了弥合高性能与人类可解释性之间的鸿沟,本文引入了 DIFFINT,这是一种创新的自编码器,其特征在于一个潜空间瓶颈,该瓶颈被结构化为一组软轴对齐的区间成员关系。DIFFINT 直接从原始数值数据中进行端到端学习,无需离散化或二值化,从而在保持深度学习表征能力的同时,维持了透明、可审计的内部结构。


🔍 Key Contributions & Methodology

  • Human-Readable Latent Space: Each latent unit corresponds to an intuitive hyper-rectangle in the feature space. An instance is encoded based on how strongly it falls inside each interval relative to other units, using reconstruction error as the final anomaly score.
  • Rigorous Inductive Biases:
  • Establishes a certified reconstruction-error lower bound for data points falling outside every active coordinate of the learned support (utilizing a Lipschitz-enforced decoder).
  • Implements a graded, empirically verified suppression mechanism tailored for cases where only a few features are abnormal.
  • Label-Free Importance Ranking: Provides a closed-form metric that ranks each (unit, feature) pair using values the model already maintains. This transforms trained intervals into auditable candidate constraints without requiring any anomaly labels.
  • 人类可读的潜空间: 每个潜单元都对应于特征空间中的一个直观超长方体。模型根据实例落入每个区间相对于其他单元的强度对其进行编码,并使用重构误差作为最终的异常得分。
  • 严谨的归纳偏置:
  • 为落在所学支持集的所有活动坐标之外的数据点建立了认证的重构误差下界(利用了施加了 Lipschitz 约束的解码器)。
  • 针对仅有少数特征异常的情况,实现了一种分级的、经实证验证的抑制机制。
  • 无标签重要性排序: 提供了一个闭式(closed-form)度量标准,使用模型本身已维护的值对每个 (unit, feature) 对进行排序。这无需任何异常标签,就能将训练好的区间转化为可审计的候选约束条件。

📊 Empirical Performance

Evaluated across 48 ADBench benchmarks against 22 baseline models under a standardized [-1, 1]-normalized protocol: * Top Performance: DIFFINT achieves the best mean rank overall on both primary metrics: * ROC-AUC: 4.10 * AUPR: 4.16 * Regime Dominance: Leads clearly among inlier-only detectors and remains highly competitive against the strongest contaminated-data detectors. * Leaderboard Standing: It stands as the only interpretable detector within the statistically-tied leading cluster of seven methods.

在标准化的 [-1, 1] 归一化协议下,针对 48 个 ADBench 基准测试 对比 22 个基线模型 进行了评估: * 顶级性能: DIFFINT 在两个主要指标上均获得了整体最佳平均排名: * ROC-AUC: 4.10 * AUPR: 4.16 * 方案统治力: 在纯正常数据(inlier-only)检测器中明显领先,并且在对抗最强污染数据检测器时保持了极强的竞争力。 * 排行榜地位: 在统计学意义上并列领先的七个方法集群中,它是唯一的可解释检测器