Donato Cerciello (UPM) presented his research on deep clustering for time-series data, focusing on learning meaningful representations and identifying patterns in temporal data without relying on labelled training examples. His work contributes to WP2 by addressing the challenges of representation learning and pattern discovery through advanced deep learning methods.

He introduced FMMVCC (Fuzzy Mamba-based Multi-View Contrastive Clustering), a framework that combines a Mamba-based encoder, contrastive learning and fuzzy clustering to integrate information from multiple views of time-series data. He presented evaluation results on univariate and multivariate benchmark datasets and discussed extending the approach to federated multi-view clustering for fall detection, addressing the analysis of distributed healthcare data.

Donato also presented explainability analyses aimed at understanding the learned clusters and identifying the temporal information relevant to the results. His work on univariate time-series clustering received the Best Student Paper Award at MLISE 2026. He additionally reported submissions on federated fall detection and the explainability of the proposed clustering approach, demonstrating how his research connects time-series representation learning with privacy-aware analysis and interpretability.

Deep_Clustering_for_Time_Series.pdf

Leonardo Schiavo (UPM) presented his research on anomaly detection in time-series data, focusing on identifying different types of anomalies, including point anomalies, contextual anomalies, anomalous sequences and irregularities in relationships between variables. His work contributes to WP2 by addressing the challenges of detecting unusual temporal patterns in complex data, particularly when anomalies involve multiple variables or extend over time.

He presented his ongoing work on TimeVQVAE-AD-M, an approach to multivariate time-series anomaly detection based on discrete representation learning and generative modelling. Building on earlier research into univariate time-series anomaly detection, he is extending the methodology to multivariate data. An important objective is to identify not only the presence of an anomaly but also the channels involved and the time intervals in which anomalous behaviour occurs.

Leonardo also discussed the use of counterfactual explanations to investigate how a time series might have looked in the absence of a detected anomaly. This direction aims to make anomaly detection results more interpretable and useful for further analysis. His presentation outlined ongoing methodological development and future research directions, including benchmarking and the investigation of federated settings.

Anomaly_Detection_for_Time_Series.pdf

Milad Jafari Barani (UNIOVI) presented his research on developing reliable and explainable AI methods for chronic disease analysis, with a particular focus on Alzheimer’s disease and longitudinal medical data. His work contributes to WP2 by investigating how temporal modelling can support the analysis of clinical trajectories while ensuring that model predictions and their explanations can be systematically evaluated.

He presented the development of recurrent neural network models, progressing from an attention-based LSTM architecture processing longitudinal clinical sequences to a dual-input stacked LSTM combining temporal data with static biomarkers. He also described a systematic framework for evaluating explanations, comprising nine sanity checks and five faithfulness tests. The experiments examined explanation stability under different types and levels of noise, including Gaussian, uniform and salt-and-pepper noise.

Milad reported results from experiments on several datasets. A proof-of-concept study using a Kaggle Alzheimer’s disease dataset was published at IEEE FLICS 2026, while experiments using the ADNI cohort, comprising 1,968 subjects, applied the full 14-test evaluation framework; the associated journal manuscript was reported as under review. An important finding, termed the “IG Paradox”, was that Integrated Gradients performed poorly on rank-based sanity checks while achieving the highest reported faithfulness among the compared methods ((\rho = 0.721), (p = 0.019)). This result highlighted that different evaluation criteria can lead to conflicting assessments of explanation quality, underlining the importance of comprehensive evaluation when developing trustworthy AI for medical applications.

Reliable__Explainable_AI_for_Chronic_Diseases.pdf