In the Artificial Intelligence research field, neurosymbolic approaches aim to combine neural networks and symbolic models to overcome their respective limitations and pave the way for more robust and explainable AI capable of reasoning. A notable example of this neurosymbolic integration is represented by the Logic Tensor Network (LTN), a framework leveraging Real Logic to build a model that maximizes the satisfiability of a set of logical facts in a knowledge base. With the goal of exploiting rule-based knowledge in graph-based tasks, this thesis proposes a framework called LTN-GCN, which merges LTNs and Graph Convolutional Neural Networks (GCNs) to address node and edge classification within graphs. The experimental results on a citation graph and a drug-drug interaction network show that LTN-GCN achieves performance levels comparable to other GCN models.
Nel campo della ricerca sull'Intelligenza Artificiale, gli approcci neurosimbolici uniscono le reti neurali e i modelli simbolici per superare le rispettive limitazioni e favorire lo sviluppo di un'AI più robusta, explainable e capace di reasoning. Un esempio notevole di questa integrazione neurosimbolica è rappresentato da Logic Tensor Network (LTN), un framework che sfrutta la Real Logic per costruire modelli che massimizzano la soddisfacibilità di un insieme di fatti logici appartenenti a una knowledge base. Con l'obiettivo di impiegare la conoscenza rule-based per risolvere task basati su grafi, questa tesi propone un framework chiamato LTN-GCN, che combina LTN e Graph Convolutional Neural Networks (GCN) per risolvere problemi di classificazione di nodi o di archi all'interno di grafi. I risultati sperimentali relativi a un citation graph e a una drug-drug interaction network mostrano che LTN-GCN raggiunge livelli di performance confrontabili con altri modelli di tipo GCN.
Un nuovo framework con Logic Tensor Networks per la risoluzione di problemi basati su grafi
SCARAGGI, VITO
2023/2024
Abstract
In the Artificial Intelligence research field, neurosymbolic approaches aim to combine neural networks and symbolic models to overcome their respective limitations and pave the way for more robust and explainable AI capable of reasoning. A notable example of this neurosymbolic integration is represented by the Logic Tensor Network (LTN), a framework leveraging Real Logic to build a model that maximizes the satisfiability of a set of logical facts in a knowledge base. With the goal of exploiting rule-based knowledge in graph-based tasks, this thesis proposes a framework called LTN-GCN, which merges LTNs and Graph Convolutional Neural Networks (GCNs) to address node and edge classification within graphs. The experimental results on a citation graph and a drug-drug interaction network show that LTN-GCN achieves performance levels comparable to other GCN models.File | Dimensione | Formato | |
---|---|---|---|
tesi_Scaraggi.pdf
embargo fino al 27/04/2026
Dimensione
3.45 MB
Formato
Adobe PDF
|
3.45 MB | Adobe PDF |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/19225