[论文解读] Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift

Stephan Rabanser, Stephan Günnemann+1 more|arXiv (Cornell University)|Oct 29, 2018

Anomaly Detection Techniques and Applications参考文献 58被引用 140

一句话总结

本论文实证比较使用降维和两样本检验的方法来检测数据集分布漂移，发现带软预测的黑箱漂移检测器在大多数漂移中表现最好，且单变量检验在实践中可以与多变量检验相匹配。它还通过领域区分分类器分析漂移的恶性程度。

ABSTRACT

We might hope that when faced with unexpected inputs, well-designed software systems would fire off warnings. Machine learning (ML) systems, however, which depend strongly on properties of their inputs (e.g. the i.i.d. assumption), tend to fail silently. This paper explores the problem of building ML systems that fail loudly, investigating methods for detecting dataset shift, identifying exemplars that most typify the shift, and quantifying shift malignancy. We focus on several datasets and various perturbations to both covariates and label distributions with varying magnitudes and fractions of data affected. Interestingly, we show that across the dataset shifts that we explore, a two-sample-testing-based approach, using pre-trained classifiers for dimensionality reduction, performs best. Moreover, we demonstrate that domain-discriminating approaches tend to be helpful for characterizing shifts qualitatively and determining if they are harmful.

研究动机与目标

• Motivate the need for detecting dataset shift to avoid silent failures in ML systems.
• Evaluate a pipeline combining dimensionality reduction with two-sample tests for shift detection.
• Identify which reduction techniques and tests perform best across diverse shifts.
• Characterize shifts qualitatively using domain-discriminating classifiers and discuss shift malignancy.

提出的方法

• Represent test data with multiple dimensionality reduction (DR) techniques including No Reduction (NoRed), PCA, Sparse Random Projection (SRP), Autoencoders (TAE, UAE), and Label Classifiers (BBSDs, BBSDh).
• Apply statistical two-sample tests: Multivariate Kernel Two-Sample Tests (Maximum Mean Discrepancy, MMD) and Univariate tests (Kolmogorov-Smirnov with Bonferroni, Chi-Squared, Binomial).
• For BBSDs and BBSDh, use classifier outputs as DR representations; for Classif, train a domain classifier to distinguish source vs target and test its significance.
• Use permutation tests for MMD, KS with Bonferroni for multiple dimensions, chi-squared for label distributions, and binomial test for domain classifier accuracy.
• Provide a mechanism to extract the most anomalous samples via domain classifier confidences when a shift is detected.
• Discuss a practical framework to assess shift malignancy by comparing model accuracy on samples most confidently labeled as target.

实验结果

研究问题

• RQ1 Can a two-sample testing framework detect distribution shift between source and target data using dimensionality-reduced representations?
• RQ2 Which dimensionality reduction techniques yield the strongest shift-detection performance across various perturbations?
• RQ3 How do univariate and multivariate two-sample tests compare in detecting high-dimensional shifts?
• RQ4 Can domain-discriminating classifiers help characterize the type and malignancy of detected shifts?
• RQ5 What is the practicality of identifying the most anomalous samples and estimating shift impact on model performance?

主要发现

• BBSDs (soft predictions) achieve strong shift-detection performance across many shifts, often outperforming other DR methods.
• Univariate testing with multiple dimensions and Bonferroni correction can match multivariate kernel tests in effectiveness.
• UAE (autoencoder) and TAE (autoencoder) perform best for multivariate testing, outperforming other DR methods in that setting.
• Domain classifier-based approaches provide qualitative insights and enable identification of the most anomalous samples, aiding shift characterization.
• The malignancy detector using domain predictions can indicate whether a shift meaningfully degrades model accuracy, with fewer labeled examples needed than full target labels.
• The paper observes that some shifts (e.g., large image rotations, significant noise) are easily detectable even with few samples, while others (e.g., small perturbations) are harder to detect.