TY - JOUR AB - We cite correspondences between dynamics in competitive markets and information theory in the objective of recovering signal from noisy information sequences. In financial markets, this objective has been examined as recovering signal on phase transitions between ordered and disordered states of agents in the market. These transitions have been indicated to denote critical points in time series of market price. Although there is a noteworthy background in information theory in the study of the dynamics of the Ising model in this manuscript, we pursue a different modeling approach. Whereas phase transitions in a multicomponent model of market states have previously been studied with numerical methods, we provide an analytical demonstration that a multicomponent model as an Ising analogue can evidence phase transitions. AU - Raseta, M.* AU - Silver, S.D.* AU - Bazarova, A. C1 - 76080 C2 - 58386 CY - 445 Hoes Lane, Piscataway, Nj 08855-4141 Usa SP - 181336-181348 TI - Phase transitions in an ising model of agent expectations in financial markets: Analytics and numerical results in one and two-dimensional network topologies. JO - IEEE Access VL - 13 PB - Ieee-inst Electrical Electronics Engineers Inc PY - 2025 SN - 2169-3536 ER - TY - JOUR AB - Cardiovascular diseases are the leading cause of global mortality, necessitating early detection and continuous monitoring for timely interventions. Smartwatches with electrocardiogram (ECG) recording capabilities enable real-time, at-home cardiac monitoring. Specific ECG characteristics can provide insights into cardiovascular diseases. The delineation of ECGs, which is the identification of fiducial points (such as onsets, offsets, and peaks), is a time-consuming task. Automated ECG delineation can enhance this process, but existing research comparing available algorithms is limited. Furthermore, to the best of our knowledge, none have addressed single-lead ECGs from smartwatches, which can be noisy and unfiltered. Thus, this study evaluates the best-performing open-source algorithm for single-lead and smartwatch ECG data. We used two public datasets (Lobachevsky University Database, QT Database) and two smartwatch datasets (SmartHeartWatch Dataset, SMART Start Dataset) including two devices (AppleWatch, Withings ScanWatch). Algorithms from three toolkits (NeuroKit, ECGKit, ECGdeli) were assessed based on the time deviation between algorithm outputs and reference annotations, sensitivity, true positives, and false negatives. Results were further evaluated against the Common Standards for Quantitative Electrocardiography (CSE) recommendations. ECGdeli outperformed the other algorithms. For QRS on- and offset, ECGkit shows comparable sensitivity, but otherwise lower scores. NeuroKit consistently shows lower sensitivity across all four data sets, however, the temporal deviation between detected point and reference was higher. Overall, sensitivity scores were higher for Apple Watch data compared to Withings ScanWatch data. This study demonstrates that segmentation algorithms are applicable to single-lead smartwatch ECG data, with ECGdeli being the most stable overall, and NeuroKit recommended for scenarios prioritizing the temporal accuracy of detected points. AU - Jaeger, K.M.* AU - Nissen, M.* AU - Flaucher, M.* AU - Gräf, L.* AU - Joanidopoulos, J.* AU - Anneken, L.* AU - Huebner, H.* AU - Goossens, C.* AU - Titzmann, A.* AU - Pontones, C.* AU - Fasching, P.A.* AU - Beckmann, M.W.* AU - Eskofier, B.M. AU - Leutheuser, H.* C1 - 72356 C2 - 56574 TI - Systematic comparison of ECG delineation algorithm performance on smartwatch data. JO - IEEE Access PY - 2024 SN - 2169-3536 ER - TY - JOUR AB - Clustering similarity measures are essential for evaluating clustering results and ensuring diversity in multiple clusterings of the same dataset. Common indices like the Mutual Information (MI) and Rand Index (RI) are biased towards smaller clusters and are often adjusted using a random permutation model. Recent advancements have standardized these measures to further correct biases, but the impact of different random models on these standardized measures has not yet been studied. In this work, we introduce equations for standardizing the MI/RI under non-permutation models, specifically focusing on a uniform model over all clusterings and a model that fixes the number of clusterings. Our results show that while standardization improves performance for the fixed number of clusters model, its benefits are limited in the more general uniform model. We validate our findings with gene expression data, highlighting the importance of choosing the right similarity metric for clustering comparison. AU - Klede, K.* AU - Altstidl, T.* AU - Zanca, D.* AU - Eskofier, B.M. C1 - 72685 C2 - 56676 CY - 445 Hoes Lane, Piscataway, Nj 08855-4141 Usa SP - 179879-179890 TI - The impact of random models on standardized clustering similarity. JO - IEEE Access VL - 12 PB - Ieee-inst Electrical Electronics Engineers Inc PY - 2024 SN - 2169-3536 ER - TY - JOUR AB - The prediction of the remaining time for business processes is a major task in predictive process monitoring (PPM). In the last years, various machine learning methods were introduced which reduced error levels steadily. However, the commonly applied metric for optimization and evaluation, the Mean Absolute Error (MAE), has limitations regarding its interpretability. In this work we introduce and evaluate the normalized Mean Absolute Error (nMAE) as an interpretable metric for model evaluation. It accounts for different kinds of label shifts, which are a special type of concept drift that can distort remaining time results. We investigate these concepts in a thorough benchmark study and use them to assess the current state of remaining time prediction for business processes. This includes the evaluation of four different baseline models, identifying the most accurate one. Furthermore, our study compares three different state-of-The-Art methods, namely XGBoost, DA-LSTM, and PGT-Net. In contrary to prior studies we find that there is no significant difference in the performance between these models. Additionally, using the nMAE as evaluation metric we find that these models do not perform reasonably well on a range of event logs. Initial ideas for this behaviour are discussed and consolidated along with other findings from the case study into a comprehensive list motivating future research directions. AU - Roider, J.* AU - Nguyen, A.* AU - Zanca, D.* AU - Eskofier, B.M. C1 - 71813 C2 - 56164 CY - 445 Hoes Lane, Piscataway, Nj 08855-4141 Usa SP - 130583-130601 TI - Assessing the performance of remaining time prediction methods for business processes. JO - IEEE Access VL - 12 PB - Ieee-inst Electrical Electronics Engineers Inc PY - 2024 SN - 2169-3536 ER - TY - JOUR AB - Federated Learning (FL) is a machine learning technique in which collaborative and distributed learning is performed, while the private data reside locally on the client. Rather than the data, only gradients are shared among all collaborative nodes with the help of a central server. To ensure the data privacy, the gradients are prone to the deformation, or the representation is perturbed before sharing, ultimately reducing the performance of the model. Recent studies show that the original data can still be recovered using latent space (i.e., gradient leakage problem) by Generative Adversarial Network and different optimization algorithms such as Bayesian and Covariance Matrix Adaptation Evolution Strategy. To address the issues of data privacy and gradient leakage, in this paper, we train deep neural networks by exploiting the blockchain-based Swarm Learning (SL) framework. In the SL scheme, instead of sharing perturbed or noisy gradients to the central server, we share the original gradients among authenticated (i.e., blockchain-based smart contract) training nodes. To demonstrate the effectiveness of the SL approach, we evaluate the proposed approach using the standard CIFAR10 and MNIST benchmark datasets and compare it with the other existing methods. AU - Madni, H.A.* AU - Umer, R.M. AU - Foresti, G.L.* C1 - 67587 C2 - 53595 CY - 445 Hoes Lane, Piscataway, Nj 08855-4141 Usa SP - 16549-16556 TI - Blockchain-based swarm learning for the mitigation of gradient leakage in federated learning. JO - IEEE Access VL - 11 PB - Ieee-inst Electrical Electronics Engineers Inc PY - 2023 SN - 2169-3536 ER - TY - JOUR AB - Pollen allergies have become one of the most wide-spread afflictions that impact quality of life. This has made the need for automatic pollen detection, classification and monitoring a very important topic. This paper introduces a new public annotated image data-set of pollen with almost 45 thousand samples obtained from an automatic instrument. In this work we apply some of the best performing convolutional neural networks architectures on the task of pollen classification as well as some fully convolutional networks optimized for image segmentation on complex microscope images. We obtain an F1 scores of 0.95 on the new data-set when the best trained model is used as a fully convolutional classifier and a class mean Intersection over Union (IoU) of 0.88 when used as an object detector. AU - Boldeanu, M.* AU - González-Alonso, M.* AU - Cucu, H.* AU - Burileanu, C.* AU - Maya-Manzano, J.M. AU - Buters, J.T.M. C1 - 65727 C2 - 52472 SP - 73675-73684 TI - Automatic pollen classification and segmentation using U-nets and synthetic Data. JO - IEEE Access VL - 10 PY - 2022 SN - 2169-3536 ER - TY - JOUR AB - In this paper we propose a distributed architecture to provide machine learning practitioners with a set of tools and cloud services that cover the whole machine learning development cycle: ranging from the models creation, training, validation and testing to the models serving as a service, sharing and publication. In such respect, the DEEP-Hybrid-DataCloud framework allows transparent access to existing e-Infrastructures, effectively exploiting distributed resources for the most compute-intensive tasks coming from the machine learning development cycle. Moreover, it provides scientists with a set of Cloud-oriented services to make their models publicly available, by adopting a serverless architecture and a DevOps approach, allowing an easy share, publish and deploy of the developed models. AU - Lopez Garcia, A.* AU - Tran, V.* AU - Alic, A.S.* AU - Caballer, M.* AU - Plasencia, I.C.* AU - Costantini, A.* AU - Dlugolinsky, S.* AU - Duma, D.C.* AU - Donvito, G.* AU - Gomes, J.* AU - Heredia Cacha, I.* AU - De Lucas, J.M.* AU - Ito, K. AU - Kozlov, V.Y.* AU - Nguyen, G.* AU - Orviz Fernandez, P.* AU - Sustr, Z.* AU - Wolniewicz, P.* AU - Antonacci, M.* AU - zu Castell, W. AU - David, M.* AU - Hardt, M.* AU - Lloret Iglesias, L.* AU - Molto, G.* AU - Plociennik, M.* C1 - 58502 C2 - 48510 SP - 18681-18692 TI - A cloud-based framework for machine learning workloads and applications. JO - IEEE Access VL - 8 PY - 2020 SN - 2169-3536 ER -