Explore a scalable and lenient two-sample test of dataset shift based on outlier scores in this conference talk from the Toronto Machine Learning Series. Learn about the D-SOS test, which focuses on low-density regions of data distribution rather than inliers. Discover how this method reduces high-dimensional data to an outlier score, separates inliers from outliers, and compares contamination rates across samples. Understand the significance of the weighted area under the receiver operating characteristic curve (WAUC) as the test statistic. Gain insights into how the D-SOS test scales with sample size and dimension, outperforming existing methods in certain scenarios. Recognize the importance of monitoring and validating data in the era of data-driven algorithms, and how the D-SOS test precedes model drift by highlighting the critical role of data quality in machine learning and AI applications.
Overview
Syllabus
Vathy M. Kamulete - D-SOS: (D)ataset (S)hift with (O)utlier (S)cores
Taught by
Toronto Machine Learning Series (TMLS)