The article discusses Iterated Relevance Matrix Analysis (IRMA), a method for identifying a linear subspace that represents class-specific information in a data set by repeatedly retraining Generalized Matrix Learning Vector Quantization (GMLVQ) classifiers. IRMA facilitates a detailed analysis of feature relevances, enables improved low-dimensional representations and visualizations of labeled data sets, and can be used for dimensionality reduction. The authors suggest that the IRMA-based class-discriminative subspace can be used to train robust classifiers with potentially improved performance.
Publication date: 23 Jan 2024
Project Page: https://arxiv.org/abs/2401.12842v1
Paper: https://arxiv.org/pdf/2401.12842
