Logo Logo
Switch Language to German
Feldhans, Robert; Wilke, Adrian; Heindorf, Stefan; Shaker, Mohammad Hossein; Hammer, Barbara; Ngonga Ngomo, Axel-Cyrille; Hüllermeier, Eyke ORCID: 0000-0002-9944-4108 (November 2021): Drift Detection in Text Data with Document Embeddings. Intelligent Data Engineering and Automated Learning – IDEAL 2021, November 25-27, 2021, Manchester, United Kingdom.
Full text not available from 'Open Access LMU'.


Collections of text documents such as product reviews and microblogs often evolve over time. In practice, however, classifiers trained on them are updated infrequently, leading to performance degradation over time. While approaches for automatic drift detection have been proposed, they were often designed for low-dimensional sensor data, and it is unclear how well they perform for state-of-the-art text classifiers based on high-dimensional document embeddings. In this paper, we empirically compare drift detectors on document embeddings on two benchmarking datasets with varying amounts of drift. Our results show that multivariate drift detectors based on the Kernel Two-Sample Test and Least-Squares Density Difference outperform univariate drift detectors based on the Kolmogorov-Smirnov Test. Moreover, our experiments show that current drift detectors perform better on smaller embedding dimensions.