Data Representation at Scale: Extracting Meaningful Knowledge from Massive Datasets

From genomic sequence analysis to real-time sustainability monitoring, and from medical decision support to neurophysiological signal processing, Big Data meaningful knowledge extraction boosts scientific discovery and societal progress. Large-scale biological sequences reveal information about living organisms, while IoT sustainability data show how socio-environmental systems work. Clinical and imaging data support surgery planning, while EEG signals provide insights into cognitive processes. From a computational perspective, these heterogeneous sources share a common issues: transforming raw, high-dimensional, and noisy observations into structured representations that are both machine learning-ready and semantically meaningful.This PhD thesis proposes novel knowledge extraction techniques from heterogeneous and large-scale data, integrating distributed Big Data infrastructures, vector and graph-based representations, and machine learning methods ranging from classical approaches to deep learning. Heterogeneous input data, time series, images, biological sequences, and text, are processed through novel frameworks for meaningful feature extraction and embedding, producing structured latent representations Z ∈ RN×d that support statistical and predictive analysis of learned features in all domains and, in the sustainability domain, a knowledge graph that encodes relational structure between entities.Four application domains are considered: Biological data, Neurophysiological data, Sustainability data, and Medical data. For Biological data, the focus is on analyzing genomic sequences and constructing compact sequence representations that combine two complementary components: a distributed Burrows–Wheeler Transform (BWT) index, providing compressed and searchable access to large sequence collections, and the BioSet2Vec framework, devoted to the extraction of compact and statistically significant k-mer dictionaries that characterize biological sequence sets. With respect to Neurophysiological data, the aim is to extract neural signatures from EEG signals as cognitive knowledge representations. In the Sustainability context, there is modeling of stakeholder relationships via knowledge graphs and real-time energy monitoring via edge–cloud analytics. Lastly, for Medical data multimodal surgical decision support in reductive mastoplasty is addressed via probabilistic fusion of clinical and meaningful morphometric representations.