@inproceedings{AmlingScheeleSlanyetal., author = {Amling, Jonas and Scheele, Stephan and Slany, Emanuel and Lang, Moritz and Schmid, Ute}, title = {Explainable AI for Mixed Data Clustering}, series = {Explainable AI for Mixed Data Clustering:Second World Conference, xAI 2024, Valletta, Malta, July 17-19, 2024, Proceedings, Part II}, booktitle = {Explainable AI for Mixed Data Clustering:Second World Conference, xAI 2024, Valletta, Malta, July 17-19, 2024, Proceedings, Part II}, publisher = {Springer Nature Switzerland}, address = {Cham}, isbn = {9783031637964}, issn = {1865-0929}, doi = {10.1007/978-3-031-63797-1_3}, pages = {42 -- 62}, abstract = {Clustering, an unsupervised machine learning approach, aims to find groups of similar instances. Mixed data clustering is of particular interest since real-life data often consists of diverse data types. The unsupervised nature of clustering emphasizes the need to understand the criteria for defining and distinguishing clusters. Current explainable AI (XAI) methods for clustering focus on intrinsically explainable clustering techniques, surrogate model-based explanations utilizing established XAI frameworks, and explanations generated from inter-instance distances. However, there exists a research gap in developing post-hoc methods that directly explain clusterings without resorting to surrogate models or requiring prior knowledge about the clustering algorithm. Addressing this gap, our work introduces a model-agnostic, entropy-based Feature Importance Score for continuous and discrete data, offering direct and comprehensible explanations by highlighting key features, deriving rules, and identifying cluster prototypes. The comparison with existing XAI frameworks like SHAP and ClAMP shows that we achieve similar fidelity and simplicity, proving that mixed data clusterings can be effectively explained solely from the distributions of the features and assigned clusters, making complex clusterings comprehensible to humans.}, language = {en} } @inproceedings{WissingScholzSalomanetal., author = {Wissing, Julio and Scholz, Teresa and Saloman, Stefan and Fargueta, Lidia and Junger, Stephan and Stefani, Alessio and Tschekalinskij, Wladimir and Scheele, Stephan and Schmid, Ute}, title = {SPECTRE: A Dataset for Spectral Reconstruction on Chip-Size Spectrometers with a Physics-Informed Augmentation Method}, series = {2024 IEEE SENSORS, 20-23 October 2024, Kobe, Japan}, booktitle = {2024 IEEE SENSORS, 20-23 October 2024, Kobe, Japan}, publisher = {IEEE}, doi = {10.1109/SENSORS60989.2024.10784898}, pages = {1 -- 4}, abstract = {Reconstructing usable spectra from an array of optical filters has been central to miniature spectrometers for many years. However, as optical sensors become more suitable for low-cost applications by exploiting sub-par filter characteristics, creating high-quality reconstructions has become increasingly challenging. There has been growing interest in data-driven reconstruction methods within the scientific community to address this issue. The main obstacle to progress in this area is the lack of publicly available and comprehensive datasets for spectral reconstruction. In our research, we propose a physics-informed data augmentation technique to significantly increase the size of datasets for spectral reconstruction and make our augmented dataset, containing 10117 samples, publicly available. Furthermore, we demonstrate the potential performance improvements achieved by training complex neural networks for spectral reconstruction using the augmented dataset. Our short benchmark indicates a performance increase of up to four times over the baseline.}, language = {en} }