Interpretable Time-series Analysis

Time-series analysis is a central task in many data-intensive domains, including industrial monitoring, environmental sensing, geoscience, finance and decision-support systems. DTALab studies time-series methods with a specific focus on interpretability, robustness and applicability to complex real-world data.

The research line focuses on four main directions:

Interpretable clustering of multivariate time series: methods that extract meaningful features from multivariate signals and produce clusters that can be inspected and understood by domain experts.
Novelty and anomaly detection: techniques for identifying new or abnormal operating conditions in industrial and monitoring scenarios, with attention to the explanation of which signals contribute to the detected change.
Forecasting of irregularly sampled time series: transformer-based models for forecasting multivariate time series affected by irregular sampling, missing values and heterogeneous temporal dependencies.
Self-supervised learning for time-series forecasting: data-centric methods that exploit decomposed representations of time series to improve forecasting models in label-scarce or complex multivariate settings.

A key result in this area is Time2Feat, an end-to-end system for clustering multivariate time series through interpretable inter-signal and intra-signal features. Time2Feat applies feature selection and dimensionality reduction to retain the most informative features and improve the interpretability of the resulting clusters.

The forecasting line includes ISTF and TED4STL. ISTF addresses forecasting of irregularly sampled multivariate time series with transformer encoders, combining temporal regularization, missing-value tracking and local/global attention mechanisms. TED4STL introduces a trend-error decomposition pipeline for self-supervised time-series learning, decomposing each series into trend and error components and evaluating whether these representations improve multivariate forecasting.

The research also includes domain-oriented applications. In industrial monitoring, autoencoder-based models are studied for novelty detection and system health monitoring. In geoscience, interpretable clustering methods have been applied to PS-InSAR time series for ground deformation detection.

This research contributes to trustworthy time-series analytics where models must not only produce accurate predictions, clusters or anomaly scores, but also expose patterns, signals and temporal behaviours that can be validated by experts.

Selected publications

Sara Pederzoli, Francesco Del Buono, Maurizio Vincini, Francesco Guerra. Trend-Error Decomposition for Self-Supervised Time Series Learning in Multivariate Forecasting Task. IEEE Access 14: 8618-8631, 2026.
Riccardo Benassi, Francesco Del Buono, Giacomo Guiduzzi, Francesco Guerra. Forecasting Irregularly Sampled Time Series with Transformer Encoders. ECML PKDD 2025, LNCS 16020: 201-217, 2026.
Claudia Masciulli, Giacomo Guiduzzi, Donato Tiano, Marta Zocchi, Francesco Guerra, Paolo Mazzanti, Gabriele Scarascia Mugnozza. Interpretable clustering of PS-InSAR time series for ground deformation detection. Computers & Geosciences 203: 105959, 2025.
Angela Bonifati, Francesco Del Buono, Francesco Guerra, Miki Lombardi, Donato Tiano. Interpretable Clustering of Multivariate Time Series with Time2Feat. Proceedings of the VLDB Endowment 16(12): 3994-3997, 2023.
Angela Bonifati, Francesco Del Buono, Francesco Guerra, Donato Tiano. Time2Feat: Learning Interpretable Representations for Multivariate Time Series Clustering. Proceedings of the VLDB Endowment 16(2): 193-201, 2022.
Francesco Del Buono, Francesca Calabrese, Andrea Baraldi, Matteo Paganelli, Francesco Guerra. Novelty Detection with Autoencoders for System Health Monitoring in Industrial Environments. Applied Sciences 12(10): 4931, 2022.

Software

For a complete list of publications, see the DBLP profile and the UNI-FIND profile.