|SCALIA GABRIELE||Cycle: XXXIII |
Section: Computer Science and Engineering
Tutor: TANCA LETIZIA
Advisor: PERNICI BARBARA Major Research topic
:Quality-aware graph-based neural networks with complex data managementAbstract:
This is an interdisciplinary thesis and it is the first of this interdisciplinary research project between information technology and chemical engineering. The project as a whole investigates the use of big data
techniques for the automatic management and the analysis of data characterizing the chemical kinetic domain.
Overall, the main contribution currently pursued by this PhD thesis is the investigation of methodologies to store, interpret and analyse complex graph-based data, with a focus on data quality awareness and deep learning techniques, using as reference example data found in the chemistry domain, but targeting domain-independent techniques.
The specificities and the boundaries of the proposed research are those which should provide the main original contributions of the proposed thesis. These stem from the chemical kinetics domain, but present challenges that are shared among many different fields. These are:
- Focus on complex/articulated data, which are common in the chemical kinetics domain, such as molecules and experimental data. These can be often interpreted as attributed graphs, where each node/edge has an influence on the others and features can be derived for the graph as a whole.
- Central role on quality management, which stems from the “open-world” assumption characterizing the target domain, involving the concept of unknown. Missing and partial data must be taken into consideration and reflected in the form of quality and uncertainty characterizing data and outputs. These need to be quantitative measure to be processed and stored.
- Usage of data-driven techniques, which stems from the need and the opportunity of re-using the increasingly large amount of available data for validation and prediction. Machine learning (and, in particular, deep learning) techniques are investigated as a complementary alternative in fields – such as chemical kinetics – where analytical models are predominant.
- Scalability and interoperability as underlying constraints. The goal is to pursue solutions that could be developed to be employed in applications in the future.