Predictive and Adaptive models for Time-varying and Heterogeneous data & systems – PATH

Research Activity

The group’s activities are organized into the following main research areas:

1. Longitudinal data analysis: it refers to a broad set of statistical and machine learning techniques used to study data collected from the same subjects repeatedly over time (i.e., Time-Series). This temporal structure makes longitudinal analysis especially valuable in contexts where predicting change, stability, or trends over time is crucial. Mainly we focus on longitudinal clustering, a flexible, data-driven approach for analyzing multivariate time-dependent data. This makes it particularly suitable for exploratory analysis, where the goal is to identify latent subpopulations based on shared temporal dynamics. A variety of methods can be developed based on classical algorithms such as k-means, fuzzy k-means, and hierarchical clustering, as well as model-based approaches like latent class growth analysis or functional data clustering. Mainly, in healthcare longitudinal clustering is crucial to analyze patient trajectories over time—such as the evolution of clinical indicators, treatment responses, or disease progression crucial to improve early diagnosis and designing targeted interventions.

2. Predictive models and Complex Networks Analysis: Building predictive models aims to explicitly capture long-range temporal dependencies, which are often key to accurate forecasting and decision support. A key challenge addressed in this research is dealing with heterogeneous and irregularly sampled data, as often encountered in real-world applications. A challenge is the prediction of complex networks evolution, including the normal and pathological changes of functional connectivity in the human brain. By focusing on the alteration of signal exchange between the major brain agglomerates, which constitute the nodes of the human connectome, a brain neurological disease can be viewed as an alteration of the network. The effect of disease as an alteration of the weights of brain connections is modelled using approaches derived from physics. The quantitative comparison between healthy brains and those affected by neurodegenerative and neuropsychiatric diseases, incorporated inside the theoretical model, allows for the development of predictive approaches aimed at predicting the development of disease.

3. Federated Learning: Federated learning approaches are required to preserve data privacy when building predictive models in distributed environments. One area of investigation will be the design of predictive maintenance models operating in a federated architecture. These models will allow different nodes to learn from local sensor data while preserving data privacy and adapting to site-specific conditions. The group will explore also federated approaches to longitudinal clustering to identify behavioural patterns or progression trajectories from temporally structured data collected across decentralised sources. In addition we will address pattern recognition in distributed systems by developing collaborative learning models that can extract knowledge from heterogeneous, heterogeneous geographically dispersed data without requiring centralisation.

4. Adaptive systems: Adaptive systems are able to operate in complex and dynamic environments by autonomously modifying their behavior in response to changes in the environment, operational conditions, user goals, and user state. These systems are characterised by the interaction between artificial agents and human users, and must be capable of perceiving, reasoning, learning, and adapting in a proactive and autonomous manner. Adaptive automated planning is a key functionality of such systems requiring different approaches including automated reasoning for plan generation, normative reasoning for managing rules, social norms, constraints, and limited resources, reinforcement learning to improve agent behavior based on experience in uncertain and partially observable environments, while also considering user preferences, cognitive architectures that integrate both deliberative and reactive reasoning for both plan generation and replanning in response to dynamic changes in the operational environment.

Goals

PATH group is dedicated to advancing the development of predictive models and adaptive systems that effectively handle dynamic, heterogeneous, and time-varying data. Our research addresses key challenges in modeling complex real-world systems across diverse domains. By integrating cutting-edge machine learning techniques with adaptive frameworks, PATH group aims to create robust, interpretable, and scalable solutions that respond intelligently to complex, dynamic, and evolving systems.

Application Fields

Healthcare, Industry 4.0, Robotics, Ambient Intelligence, Active and Assistive Living.