Student opportunities

The increasing use of wireless communication systems makes radio‑frequency (RF) channels a critical target for malicious intrusions and anomalous activities. This thesis aims to investigate deep learning–based approaches for the detection of intrusions in RF communication systems, leveraging raw or processed RF signal representations. A key objective is to integrate Explainable Artificial Intelligence (XAI) techniques to improve the interpretability of detection models, enabling insight into the signal features and patterns that drive classification decisions. The work will evaluate different neural architectures, assess detection performance under varying channel conditions, and analyze the trade‑off between accuracy and explainability. The final goal is to enhance trust, robustness, and transparency in AI‑based RF security systems. Expected activities:

• Review of deep learning approaches and XAI techniques for RF intrusion detection.
• Preparation and processing of RF signal datasets under varying channel conditions.
• Implementation of neural models for intrusion detection with explainability methods.
• Evaluation of detection performance, robustness, and interpretability trade‑offs.

THESIS BSc/MSc
Rule-based decision methods are becoming crucial in modern process automation system since they allow expert users to understand the mechanism used by the machine to take decisions. Moreover, integration to manual rules (discovered by humans) and automatic rules (discovered by the machine) is a key point for improving automatic decision processes. To this aim, it is essential to transform the space of the rules into a metric space, in order to allow operations such as distance computation.
The student will:

• Study a way to evaluate the distance between sets of rules
• Implement these definitions
• Design a case study to evaluate the effectiveness of this approach
• Perform the tests and evaluate the results

Motion analysis plays a central role in the assessment and personalization of physical rehabilitation programs. This thesis focuses on the application of Explainable Artificial Intelligence (XAI) techniques to the analysis of motion data acquired from rehabilitation patients using wearable sensors or motion capture systems. The main objective is to develop machine learning models capable of characterizing movement quality, detecting abnormal or compensatory patterns, and supporting clinical evaluation, while providing human‑interpretable explanations of model decisions. Particular attention is given to aligning AI outputs with clinically meaningful motion features to foster trust and usability in healthcare contexts. The proposed work aims to contribute to more transparent, data‑driven rehabilitation monitoring and decision support systems. Expected activities:

• Review of XAI techniques and machine learning methods for motion analysis in rehabilitation
• Preparation and preprocessing of motion data from wearable sensors or motion capture systems
• Implementation of models for movement quality assessment and abnormal pattern detection with explainability
• Evaluation of results focusing on interpretability and clinical relevance

Deep learning–based computer vision models achieve high performance across tasks such as image classification, object detection, and segmentation, but often lack transparency in their decision-making processes. This thesis investigates the use of Large Language Models (LLMs) as a complementary tool for generating human-readable explanations of visual model outputs. The objective is to bridge the gap between complex visual predictions and interpretable descriptions by leveraging multimodal reasoning and natural language generation.

The work explores methods to combine vision model outputs (e.g., feature maps, attention scores, or predicted labels) with LLM-based explanation frameworks, enabling clear and meaningful justifications of model decisions. Particular attention is given to the fidelity, consistency, and usefulness of generated explanations, as well as their alignment with human understanding. The approach will be evaluated across different computer vision tasks and datasets, assessing both quantitative performance and qualitative interpretability.

The expected outcome is to demonstrate how LLMs can enhance the transparency and trustworthiness of computer vision systems, supporting their adoption in critical domains such as healthcare, autonomous systems, and industrial inspection.

Expected Activities:

• Review of explainability methods for computer vision and LLM-based approaches to explanation.
• Integration of vision model outputs with LLMs for generating natural language explanations.
• Implementation and evaluation across selected vision tasks and datasets.
• Analysis of explanation quality, consistency, and usefulness for end users.

This internship focuses on the integration and application of time-series feature extraction techniques within the Rulex Platform, leveraging the open-source TSFEL library (Time Series Feature Extraction Library) or other similar tools. The candidate will contribute to the design and implementation of modules that enable automated extraction, selection, and management of meaningful features from temporal signals, ensuring seamless integration with the platform’s architecture and workflows.

In parallel, the work will explore a domain-specific use case, with particular attention to healthcare applications. This may include the analysis of biomedical signals (e.g., physiological monitoring data) to support tasks such as classification, anomaly detection, or decision support.

The candidate will evaluate how extracted features can enhance model interpretability and performance, potentially combining them with advanced machine learning or AI approaches. The expected outcome includes both a robust software integration and an experimental assessment demonstrating the value of feature-based representations in real-world scenarios, especially in sensitive and high-impact domains such as healthcare.

Expected activities:

• Review of time‑series feature extraction techniques and integration within the Rulex Platform using TSFEL or other similar tools.
• Design and implementation of modules for automated feature extraction, selection, and management.
• Application to healthcare time‑series data (e.g., biomedical signals) for classification or anomaly detection.
• Evaluation of feature impact on model performance and interpretability in real‑world use cases.

Foundation models for tabular data are emerging as a promising paradigm for improving predictive performance and generalization across structured datasets, with potential applications in finance, healthcare, and industrial analytics. This thesis aims to conduct a comprehensive benchmarking study of state-of-the-art foundation models tailored for tabular data, including transformer-based architectures, pretraining strategies, and hybrid approaches that integrate deep learning with traditional methods. The objective is to systematically evaluate their performance across multiple dataset variants, considering differences in size, feature distribution, missing values, and domain characteristics.

Benchmarking experiments will be designed to assess key aspects such as predictive accuracy, robustness to dataset shifts, scalability, and computational efficiency. A diverse suite of tabular datasets—both real-world and synthetically perturbed versions—will be used to highlight how model performance varies under controlled modifications of data properties. Particular attention will be given to understanding the impact of pretraining and transfer learning capabilities in low-data or domain adaptation scenarios.

The results are intended to provide practical and methodological insights into the strengths and limitations of foundation models for tabular data, offering guidance for researchers and practitioners in selecting appropriate modeling approaches for structured data tasks.

Expected activities:

• Review of state‑of‑the‑art foundation models and benchmarking methods for tabular data.
• Preparation of diverse datasets and generation of controlled variants (e.g., noise, missing values).
• Implementation and execution of benchmarking experiments across selected models.
• Analysis and comparison of results to derive practical guidelines and insights.

THESIS MSc
Among the pillar applications in supervised learning, we have classification, for which the target variable can assume a finite set of unordered values and regression, for which the target variable can assume an infinite set of ordered values. Yet, an in-between, practically very relevant case exists: ordinal regression. In this case, the target variable can assume a finite set of values, but these values are somehow ordered. For example, different degrees of risk. The candidate will explore state-of-the-art techniques in ordinal regression, having the chance also to study and experiment the adaptation of the Logic Learning Machine algorithm [1], currently available to address classification and regression problems, also to this case.

[1] Muselli M, & Ferrari E. (2011) Coupling logical analysis of data and shadow clustering for partially defined positive Boolean function reconstruction. IEEE Transaction on Knowledge and Data Engineering

Health monitoring systems increasingly rely on continuous data streams collected from wearable and ambient sensors. This thesis focuses on the application of machine learning methods to the analysis of time series data from health monitoring sensors, with the objective of extracting meaningful patterns related to physiological states, anomalies, or long‑term trends. The work explores techniques for signal preprocessing, feature extraction, modeling, and anomaly or change detection. Emphasis is placed on robustness to noise, missing data, and inter‑subject variability. The ultimate goal is to support early detection, monitoring, and decision‑making in healthcare through effective analysis of multivariate sensor time series. Expected activities:

• Review of machine learning methods for time series analysis in health monitoring.
• Preparation and preprocessing of multivariate sensor data (e.g., noise handling, missing data).
• Implementation of models for pattern extraction, anomaly detection, and trend analysis.
• Evaluation of model robustness and interpretation of results for healthcare applications.

Ensuring the resilience and safety of critical buildings requires continuous monitoring of structural behavior under operational and extreme conditions. This thesis addresses Structural Health Monitoring (SHM) through data‑driven anomaly detection techniques applied to sensor measurements such as vibrations, strains, or accelerations. The primary objective is to design and evaluate machine learning models capable of identifying early signs of structural degradation or damage without reliance on extensive labeled data. Explainable AI (XAI) methods are integrated to provide insight into detected anomalies, supporting engineers in understanding the underlying physical phenomena. The outcome of this work is intended to improve the reliability, interpretability, and practical adoption of AI‑based SHM systems for critical infrastructure. Expected activities:

• Review of data‑driven anomaly detection and XAI methods for Structural Health Monitoring.
• Preparation and preprocessing of multivariate sensor data (e.g., vibration, strain signals).
• Implementation of models for anomaly detection and explainability analysis.
• Evaluation of model reliability, interpretability, and effectiveness for SHM applications.

Efficient transport optimization is a fundamental problem in modern logistics networks, with significant impact on cost, sustainability, and service quality. This thesis aims to perform a comparative evaluation of state‑of‑the‑art routing and optimization libraries used to solve vehicle routing and transport planning problems. The objective is to analyze differences in performance, scalability, flexibility, and ease of integration across multiple real‑world scenarios and constraints. Benchmarking experiments will be conducted on representative logistics use cases, highlighting strengths and limitations of each tool. The results are intended to provide practical guidance for researchers and practitioners in selecting appropriate routing solutions for logistics applications. Expected activities:

• Review of state‑of‑the‑art routing and optimization libraries for transport problems.
• Setup of benchmark scenarios and implementation of representative logistics use cases.
• Execution of comparative experiments evaluating performance, scalability, and flexibility.
• Analysis of results to identify strengths, limitations, and practical recommendations.

THESIS MSc
Chess, known for its rich history and strategic complexity, has long been a subject of fascination for enthusiasts and scholars alike. To gain insights into the performance of chess players across time, historical data can serve as a valuable resource. The aim of this project is to analyze historical chess player performance data and uncover trends, patterns, and key milestones in the world of chess.

The outcome of this research project will be a comprehensive analysis of chess player performance throughout history. This analysis will include:

• Identification of influential chess players who made significant contributions to the game.
• Evaluation of player performance metrics, such as Elo ratings, tournament victories, and match outcomes.
• Examination of the evolution of chess strategies and opening moves.
• Insights into how technological advancements, such as computer analysis, have impacted chess.

THESIS BSc
A distributed system in cloud applications is a set of different containers which runs different codes in a correlated way, with a high level of internal communication between the parts. We used Microservices are referred to as containers exploiting a particular operation inside our system. The main used system of communication in microservices distributed system is the AMQP protocol, a broker mediated protocol that allows a high level of control about message retaining, queueing and exchange. The main AMQP broker used in many commercial and open-source application is RabbitMQ. Scope of the thesis is to study this type of communication in a complex microservice structure where many types of different queues and exchange are present and where a high level of security and performance is strictly required. After a benchmark testing phase the broker configuration will be optimized to speed up the whole system, using the most advanced data analysis technique available on the market.

THESIS BSc
In many Data Analysis applications not homogeneous data need to be treated. Data spans from number to images, from text to geographical coordinates and for each of these types of particular operations need to be implemented and offered to Data Scientists. One of the most difficult challenges in Data Analysis is data graphical representation: data are mostly visualised through a table grid, sometimes ill represented in a grid cell. The scope of this thesis is to implement a general approach to store any form of derived data (images, geographical coordinates, credit card number and so on) in a column table maintaining a fast and clear visualization. Users should be able to perform specific operations owned by the derived type (length between two geographical quantity, opacity filter for images and so on).

THESIS MSc
The most important feature of a distributed cloud application is the balance between its internal components. CPU and RAM are provided in a distributed form and the ability of the software itself to control the load between the various parts is the key feature of this type of architecture. To reach this goal, an estimate (given the operation) of the necessary CPU and RAM to complete it represents a primary input. However, in many complex scenarios a reliable estimate of such quantity is far from being trivial. The scope of this thesis project is to try to apply Machine Learning technique to this field to construct a Load Balancer Application. The Load Balancer Application will be able to learn for the precedent cases and apply autoregression techniques to better estimate the effort of any operation on the architecture itself.

THESIS MSc
Databases are the most used type of storage in data analysis, in cloud application and in software in general. They represent the most used storage class in almost all the possible fields. Databases are divided in SQL and NOSQL macro types. SQL Databases are more optimized in treating structure data but they depend on SQL syntax and they are far from being general: each SQL database distribution has its SQL syntax, its behaviour and its performance. Moreover, conflicts management (read/read, read/write, and write/write parallel access) is not treated correctly in many used distributions, leading to unexpected results in case of high parallelism as in any cloud application. For this reason, in recent years NOSQL database has grown up to fill this lack of old SQL database. NOSQL database relies on different file system logic. They allow not structured data to be stored more efficiently; they correctly treat conflict management and allow to perform simple query with high performance due to index replication. On the contrary however, they are extremely slow in fetching high quantity of data with respect to a standard SQL database, preventing them to be used in big data analysis cases. The scope of this thesis project is to implement a NOSQL database with some innovative ideas to speed up the fetching of data reducing as much as possible computation time and memory consumption (at least for simple single value queries). A comparison with standard SQL and NOSQL distribution will also be performed.

THESIS MSc
In machine learning, single and multivariate linear regression represent two of the most used algorithms to approximate a set of data with a fitting function. The searched function is a linear function in the single case, or a polynomial function in the latter. However, both the solutions return a continuous function which in some cases could be inaccurate, especially to treat fast changing process. The scope of this thesis project is to implement an innovative approach in which regression is applied searching for a piecewise linear function, with a fixed number of breakpoints. Later the behaviour of the piecewise approximation will be compared with single and multivariate solution and with other ML regression system as LLM, SVM and Neural Network.

THESIS MSc
Many decisions, both individual (trading) and organizational (resource allocation) are driven by a careful analysis of balance sheet data. Together with classical techniques, which involve the computation of some well-known indices, also machine learning approaches, which allow a richer insight on the situation, are growing in popularity. The candidate will investigate available ML balance sheet analysis techniques and will have the chance to develop an analysis pipeline of this kind without the need to write code, using the visual programming techniques of Rulex Platform.

THESIS MSc
Financial markets are not only driven by quantitative data but also heavily influenced by qualitative information. NLP, a branch of artificial intelligence, specializes in understanding and processing human language, making it a valuable tool for extracting knowledge from textual financial sources. The integration of NLP techniques can offer a competitive edge by identifying market-moving events, sentiment shifts, and hidden patterns within textual financial data.

The thesis activities will include:

• Collection of a proper set of data including e.g. financial news, earning reports, social media content...
• Implementation of NLP models for the analysis of the textual data.
• Analysis of sentiment scores to measure market sentiment about news and events.
• Perform predictive analysis using NLP-derived features.
• Interpret research findings and evaluate the potential impact of NLP technologies.