The article presents STYLIST, a new method for novelty detection, which is the identification of meaningful deviations from established data distributions. It separates semantic changes, which are relevant to the task, from style changes, which are not. STYLIST focuses on dropping environment-biased features and uses a scoring system based on feature distribution distances. The selection process removes features responsible for spurious correlations, improving novelty detection performance. The method was tested on multiple datasets, proving that the selection mechanism enhances novelty detection algorithms.
Publication date: 5 Oct 2023
Project Page: https://arxiv.org/abs/2310.03738v1
Paper: https://arxiv.org/pdf/2310.03738
