Tanir Ozcelebi

In this study, a structured and methodological evaluation approach for eXplainable Artificial Intelligence (XAI) methods in medical image classification is proposed and implemented using LIME and SHAP explanations for chest X-ray interpretations. The evaluation framework integrates two critical perspectives: predictive model-centered and human-centered evaluations. Predictive model-centered evaluations examine the explanations’ ability to reflect changes in input and output data and the internal model structure. Human-centered evaluations, conducted with 97 medical experts, assess trust, confidence, and agreements with AI’s indicative and contra-indicative reasoning as well as their changes before and after provision of explainability. Key findings of our study include explanation of sensitivity of LIME and SHAP to model changes, their effectiveness in identifying critical features, and SHAP’s significant impact on diagnosis changes. Our results show that both LIME and SHAP negatively affected contra-indicative agreement. Case-based analysis revealed AI explanations reinforce trust and agreement when participant’s initial diagnoses are correct. In these cases, SHAP effectively facilitated correct diagnostic changes. This study establishes a benchmark for future research in XAI for medical image analysis, providing a robust foundation for evaluating and comparing different XAI methods.

Explainable Artificial Intelligence (XAI) is essential for ensuring transparency and trust in healthcare. However, a systematic evaluation of XAI methods for tabular healthcare data is lacking. In this paper, we present a comprehensive benchmarking framework designed to evaluate the quality of explanations produced by local and global explainable AI (XAI) methods applied to tabular healthcare data. The framework integrates predictive model-centered evaluation, which assesses how well explanations capture model behavior, and human-centered evaluation, which measures alignment between explanations and expert-defined clinical reasoning. We apply this framework to a stroke prediction dataset and evaluate it with six predictive models (i.e., interpretable models, tree-based ensemble algorithms, and kernel-based methods) representing a wide range from interpretable to black-box models. These models are explained using nine different local and global XAI methods covering both model-agnostic and model-specific techniques to provide diverse methodological coverage. The benchmark results provide practical insights into the strengths and limitations of different XAI methods across model types and explanation scopes, offering a structured guide for selecting reliable and clinically meaningful XAI method–predictive model combinations for healthcare decision support systems.

Federico Raspanti, Tanir Ozcelebi, Mike Holenderski. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track). 2025.

Federated Learning (FL) is a distributed machine learning paradigm that enables learning models from decentralized private datasets where the labeling effort is entrusted to the clients. While most existing FL approaches assume high-quality labels are readily available on users’ devices, in reality, label noise can naturally occur in FL and is closely related to clients’ characteristics. Due to scarcity of available data and significant label noise variations among clients in FL, existing state-of-the-art centralized approaches exhibit unsatisfactory performance, whereas prior FL studies rely on excessive on-device computational schemes or additional clean data available on the server. We propose FedLN , a framework to deal with label noise across different FL training stages, namely FL initialization, on-device model training, and server model aggregation, able to accommodate the diverse computational capabilities of devices in an FL system. Specifically, FedLN computes per-client noise level estimation in a single federated round and improves the models’ performance by either correcting or mitigating the effect of noisy samples. Our evaluation on various publicly available vision and audio datasets demonstrates a 22% improvement on average compared to other existing methods for a label noise level of 60%. We further validate the efficiency of FedLN in human-annotated real-world noisy datasets and report a 4.8% increase on average in models’ recognition performance, highlighting that FedLN can be useful for improving FL services provided to everyday users.

Unobtrusive sensing (device-free sensing) aims to embed sensing into our daily lives. This is achievable by re-purposing communication technologies already used in our environments. Wireless Fidelity (Wi-Fi) sensing, using Channel State Information (CSI) measurement data, seems to be a perfect fit for this purpose since Wi-Fi networks are already omnipresent. However, a big challenge in this regard is CSI data being sensitive to ‘domain factors’ such as the position and orientation of a subject performing an activity or gesture. Due to these factors, CSI signal disturbances vary, causing domain shifts. Shifts lead to the lack of inference generalization, i.e., the model does not always perform well on unseen data during testing. We present a domain factor-independent feature-extraction pipeline called ‘mini-batch alignment’. Mini-batch alignment steers a feature-extraction model’s training process such that it is unable to separate intermediate feature-probability density functions of input data batches seen previously from the current input data batch. By means of this steering technique, we hypothesize that mini-batch alignment (i) absolves the need for providing a domain label, (ii) reduces pipeline re-building and re-training likelihood when encountering latent domain factors, and (iii) absolves the need for extra model storage and training time. We test this hypothesis via a vast number of performance-evaluation experiments. The experiments involve both one- and two-domain-factor leave-out cross-validation, two open-source gesture-recognition datasets called SignFi and Widar3, two pre-processed input types called Doppler Frequency Spectrum (DFS) and Gramian Angular Difference Field (GADF), and several existing domain-shift mitigation techniques. We show that mini-batch alignment performs on a par with other domain-shift mitigation techniques in both position and orientation one-domain leave-out cross-validation using the Widar3 dataset and DFS as input type. When considering a memory-complexity-reduced version of the GADF as input type, mini-batch alignment shows hints of recuperating performance regarding a standard baseline model to the extent that no additional performance due to weight steering is lost in both one-domain-factor leave-out and two-orientation-domain-factor leave-out cross-validation scenarios. However, this is not enough evidence that the mini-batch alignment hypothesis is valid. We identified pitfalls leading up to the hypothesis invalidation: (i) lack of good-quality benchmark datasets, (ii) invalid probability distribution assumptions, and (iii) non-linear distribution scaling issues.

Federated Learning (FL) is a distributed machine learning paradigm that enables learning models from decentralized private datasets, where the labeling effort is entrusted to the clients. While most existing FL approaches assume high-quality labels are readily available on users' devices; in reality, label noise can naturally occur in FL and is closely related to clients' characteristics. Due to scarcity of available data and significant label noise variations among clients in FL, existing state-of-the-art centralized approaches exhibit unsatisfactory performance, while prior FL studies rely on excessive on-device computational schemes or additional clean data available on server. Here, we propose FedLN, a framework to deal with label noise across different FL training stages; namely, FL initialization, on-device model training, and server model aggregation, able to accommodate the diverse computational capabilities of devices in a FL system. Specifically, FedLN computes per-client noise-level estimation in a single federated round and improves the models' performance by either correcting or mitigating the effect of noisy samples. Our evaluation on various publicly available vision and audio datasets demonstrate a 22% improvement on average compared to other existing methods for a label noise level of 60%. We further validate the efficiency of FedLN in human-annotated real-world noisy datasets and report a 4.8% increase on average in models' recognition performance, highlighting that~\method~can be useful for improving FL services provided to everyday users.