Unsupervised Methods for High-dimensional data patterns
Date:
This presentation was delivered to the ISG (Information System Group) group at CISUC. The main content addressed the exploration and development of unsupervised methods aimed at discovering patterns in high-dimensional data. The problem was contextualized by contrasting Small Data and Big Data, highlighting that smaller-scale data, despite quantitative challenges, offers greater explainability and focuses on “why” events happen.
Topics covered:
- Challenges of High Dimensionality: Exploration of the Hughes Phenomenon, illustrating how an excess of features can degrade the performance of machine learning models in complex domains such as bioinformatics, computer vision, and natural language processing (NLP).
- Unsupervised Feature Selection (UFS): Detailed overview of the KNMFS method (published in Expert Systems with Applications, 2022). This method performs feature selection through the iterative factorization of a similarity graph and modularity-focused clustering.
- Deep Subspace Clustering: Introduction to the new DCSSC (Deep Contrastive Variational Subspace Spectral Clustering) method, accepted for publication in Neurocomputing (2025). This method builds a robust representation by extracting relational information from a K-Nearest Neighbors (KNN) graph and employs a contrastive variational autoencoder (cVAE) combined with spectral clustering, outperforming both shallow and deep approaches across multiple scenarios.
- Future and Ongoing Work (GNN and TDA):
- Development of deep topological graph clustering methods based on the Mapper algorithm principle.
- Application of Graph Neural Networks (GNNs) in the computational chemistry domain, specifically analyzing SMILES molecular representations (such as the ReLeaSE dataset) for biological and chemical metrics.
- Construction of 3D Scene Graphs for spatial perception and semantic relationship modeling in object detection.