Graph Anomaly Detection with GNNs: introduction and applications in the predictive maintenance of industrial systems

• Giovanni Zurlo (Istituto Nazionale di Fisica Nucleare)

In mission-critical industries, early anomaly detection can be the difference between seamless operations and costly downtime. Graph Neural Networks (GNNs) are emerging as a powerful tool for predictive maintenance, offering unparalleled understanding of data from complex, interconnected and distributed systems of components. This session will introduce the fundamentals and real-world applications of GNNs in anomaly detection, enabling participants to gain insights into the potential benefits of core GNN techniques tailored for anomaly detection, along with current challenges and limitations. Two case studies from the industrial and IT sectors will be discussed to highlight the versatility of this approach: one focusing on sensor data from ENI S.p.A rotating machinery, and another on signals from the electrical and mechanical systems of the INFN CNAF Tier-1 data center. Join us to explore this promising field of research, engage and share ideas on new potential applications of GNNs.

Further analysis on the thesis of Megi Ceka titled From text to insights: Leveraging Topic Modelling to Explore Climate Change’s Impact on Cultural Heritage Literature

• Laura Verdesca

This study extends previous work on systematic review of the literature on the impact of climate change on cultural heritage. Our objective is to develop a robust classifier capable of categorizing new research publications into distinct thematic topics. We began by using six different topic modeling techniques—Latent Dirichlet Allocation (LDA), Latent Semantic Indexing (LSI), Hierarchical Dirichlet Process (HDP), Non-Negative Matrix Factorization (NMF), Structural Topic Modeling (STM), and Correlated Topic Modeling (CTM)—to generate topic probability distributions for each document. These topic probabilities served as input features for various classifiers. Additionally, we incorporated BERT embeddings to capture nuanced semantic information from abstracts and leverage large language models for improved topic representation. To further enhance performance, we developed ensemble methods that integrate multiple classifiers, achieving higher classification accuracy. Finally, we applied the optimized model to classify an additional dataset of 259 new papers, enabling a refined categorization of climate change’s impact on cultural heritage. We extended our work by applying the same topic model techniques to a new theme: over-tourism, to further investigate possible threats to cultural heritage.