A crucial phase in the knowledge discovery process is avoiding spurious results caused by noise or random data fluctuations. To address this, a standard approach is to evaluate the statistical significance of data mining outcomes. In network analysis, this is typically achieved by comparing observed findings against random networks drawn from a graph null model. A prominent application, and the main motivation for our work, is the study of structural polarization in online social networks. Observational studies in this domain often rely on rich interaction data, such as topic-specific collections of tweets, which can be naturally represented as vertex-colored multigraphs. Despite the critical role of this color information, traditional graph null models typically preserve only basic topological characteristics, such as the degree sequence. In this work, we consider an ensemble of metrics grounded in fundamental social network measures, such as transitivity, distances, and connectivity, that capture complex structural patterns in polarized colored networks that existing null models do not preserve explicitly. We use these metrics to systematically compare the Configuration Model, Polaris, and Sirius, aiming to identify which structural features of polarized networks remain poorly captured. Our results indicate that, while color-aware models like Polaris and Sirius better preserve properties related to assortativity than the Configuration Model, none of them fully reproduce the patterns observed in real-world polarized networks.
Una fase cruciale nel processo di knowledge discovery consiste nell'escludere risultati spuri dovuti a rumore o a fluttuazioni casuali dei dati. Per ovviare a questo problema, la prassi standard richiede di valutare la significatività statistica dei risultati di data mining. Nell'ambito della network analysis, questa validazione si ottiene tipicamente confrontando i dati osservati con reti casuali generate tramite un graph null model. Un ambito applicativo di grande interesse, che costituisce la motivazione principale di questo lavoro, è lo studio della polarizzazione strutturale nei social network online. Gli studi osservazionali in questo settore si basano spesso su vaste moli di dati, come raccolte di tweet su un tema specifico, che possono essere modellati efficacemente attraverso multigrafi a nodi colorati. Sebbene l'informazione sul colore giochi un ruolo fondamentale, i null model tradizionali per i grafi preservano soltanto le caratteristiche topologiche di base, come la degree sequence. In questo lavoro consideriamo un set di metriche, basate su misure classiche della network analysis, quali transitività, distanze e connettività, molto utili per catturare pattern strutturali complessi in reti colorate e polarizzate, che i modelli attuali non preservano esplicitamente. La tesi presenta un esteso confronto sperimentale tra il Configuration Model, Polaris e Sirius, con l'obiettivo di individuare quali caratteristiche strutturali delle reti affette da polarizzazione non vengano adeguatamente catturate da questi modelli di riferimento. I nostri risultati mostrano che, sebbene i modelli che integrano l'informazione sul colore, come Polaris e Sirius, preservino esplicitamente le proprietà di assortatività, a differenza del Configuration Model, nessuno di essi è in grado di riprodurre appieno i pattern osservati nelle vere reti polarizzate.
Analisi delle caratteristiche dei modelli nulli per reti colorate
KOSUMOVIC, TONI
2025/2026
Abstract
A crucial phase in the knowledge discovery process is avoiding spurious results caused by noise or random data fluctuations. To address this, a standard approach is to evaluate the statistical significance of data mining outcomes. In network analysis, this is typically achieved by comparing observed findings against random networks drawn from a graph null model. A prominent application, and the main motivation for our work, is the study of structural polarization in online social networks. Observational studies in this domain often rely on rich interaction data, such as topic-specific collections of tweets, which can be naturally represented as vertex-colored multigraphs. Despite the critical role of this color information, traditional graph null models typically preserve only basic topological characteristics, such as the degree sequence. In this work, we consider an ensemble of metrics grounded in fundamental social network measures, such as transitivity, distances, and connectivity, that capture complex structural patterns in polarized colored networks that existing null models do not preserve explicitly. We use these metrics to systematically compare the Configuration Model, Polaris, and Sirius, aiming to identify which structural features of polarized networks remain poorly captured. Our results indicate that, while color-aware models like Polaris and Sirius better preserve properties related to assortativity than the Configuration Model, none of them fully reproduce the patterns observed in real-world polarized networks.| File | Dimensione | Formato | |
|---|---|---|---|
|
Kosumovic_Toni.pdf
embargo fino al 22/09/2027
Dimensione
1.32 MB
Formato
Adobe PDF
|
1.32 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/114217