Neural Embeddings for Dimensionality Reduction of Complex Topology Feature Spaces

This study focuses on the main role of neural embeddings and their related design and optimization in the context of Artificial Intelligence (AI), particularly in the field of Deep Learning and Explainable AI (XAI). It explores how neural embeddings of data characterized by complex topology are crucial to address challenges and developments in the areas of data dimensionality reduction and network prediction analysis.In this thesis work, two independent but connected investigations were carried out to investigate the effect of neural encoding generated by the network for the target task in the case of data with a graph structure.The first project involved the study, design, and analysis of neural embeddings of synthetic polymers through the development of two Graph Variational Autoencoder neural networks. The goal is generating new polymers that incorporate additional structural information specific to the compounds, such as stoichiometry and chain architecture.These results were analyzed through several evaluation metrics that compare the two models created and highlight weaknesses and strengths of both approaches.A qualitative investigation of the latent space of the network highlighted how different neural embeddings created by the networks encode different information depending on the decoder model trained for generation, confirming and justifying the results obtained.In the second work, a graph neural network capable of predicting the bioactivity of molecules toward specific proteins was developed, employing neural embeddings to condense the totality of chemical information of the input data. Next, a hierarchical XAI methodology was devised to obtain additional interpretability information on molecular moieties that are relevant for the prediction, thus helping to clarify the model's decision-making process. The results obtained through explainability contribute to a deeper understanding of the data and the underlying problem.Through these studies, the importance of neural embedding design and optimization in the case of data and features with complex topology is highlighted, showing how deep neural networks, downstream of perfectly conducted training, embed all the information needed for the objective task in an encoded representation.