DataSLO

Metaheuristics alongside Neural Models for Combinatorial Optimization

(held by Josu Ceberio)

Metaheuristic algorithms have set the standard in the field of combinatorial optimization for decades. However, in recent years, due to the deep learning revolution, a large number of works have attempted to address combinatorial problems. In fact, one of the most promising lines in the research of new algorithms lies in proposing algorithms that combine ideas from both worlds.

In this course, we will begin with the fundamentals of combinatorial optimization and a review of the most recognized metaheuristic algorithms, to delve into the most current neural proposals.

Heuristics for Continuous Optimisation – Design, Replicability & Behaviour

(held by Fabio Caraffini)

Optimisation tools are essential in modern computing, with heuristic algorithms widely used for their adaptability and reliance on fitness feedback. However, the complex dynamics of these algorithms remain largely unexplored, partly because studies often emphasise only the final outcomes, neglecting valuable data produced throughout their execution.

The recent proliferation of hard-to-replicate, metaphor-based algorithms adds further complexity. This short course will cover foundational heuristics for continuous optimisation, examine modular frameworks for testing various configurations, and highlight the importance of principles enabling thorough analysis of algorithmic processes and dynamics.

We will also address methods for enhancing results replicability and for benchmarking algorithmic behaviours, including the identification of structural biases.

Mastering the Data Science Pipeline: from Acquisition to Insights

(held by Paolo Mengoni)

This course offers a comprehensive overview of the data science pipeline, emphasizing data acquisition, cleaning, transformation, and knowledge discovery through data mining tools.

Participants will engage in hands-on learning, applying theoretical concepts to practical case studies using Python programming. Through collaborative projects and real-world datasets, students will develop the ability to extract meaningful insights and enhance their problem-solving skills.

This short course will equip the students with the foundational knowledge necessary to tackle more complex data science challenges in future studies.

AnyToVec – Learning Vector Representations

(held by Alfredo Milani)

This lecture explores innovative approaches to embedding complex relationships into dense vector representations, a foundational component in current machine learning and data analysis, with data-driven application across varied domains.

From the Word2Vec, which transforms words into semantic vectors based on context, and gave recent rise to Large Language Models systems, we will explore methods like Node2Vec, Graph2Vec, and other similar frameworks that map the relationships found in various data forms, and how these models support sophisticated data interpretations across linguistic, biological, social, and other structured domains.

By embedding diverse data points into a common vector space, these techniques facilitate applications from recommendation systems to predictive modeling, where learning the "meaning" or "relationship" of data points within a network or sequence is crucial.

Causal Discovery from Time Series

(held by Maurizio Porfiri)

Discovering cause-and-effect relationships between coupled dynamical systems from time-series is a critical problem in data and network science. In this short course, students will learn how to tackle this problem through the information-theoretic concept of transfer entropy.

The first half of the short course will include a primer on information theory, statistical analysis, data pre-processing, and hypothesis-testing. Theoretical insights will be put in practice in the second half of the course, when students will create their own computer code for causal discovery and apply it to the study of real datasets.