Browsing by Issue Date, starting with "2025-11-27"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- User profiling with feature selection and explainability: essays on three case studies across different domainsPublication . Freitas, Diogo Nuno Teixeira; Teixeira Freitas, Diogo Nuno; Dias, Fernando Manuel Rosmaninho Morgado Ferrão; Fermé, Eduardo LeopoldoUser profiling is the process of constructing a structured representation of the user within a system. This representation includes information such as preferences, behaviors, and characteristics. Based on the profile, the system can recommend services and products or, in this work, suggest actions. Machine learning methods are commonly used to this end, as they can identify complex patterns among large numbers of attributes. However, not all attributes are relevant. High-dimensional datasets often contain irrelevant, redundant, or noisy features that obscure valuable patterns and reduce model accuracy. To address this, dimensionality reduction techniques—particularly feature selection—are essential. Equally important is the ability to explain a model’s output, since understanding why a model produces a given outcome builds trust and clarifies which steps can change an undesirable situation. This thesis applies feature selection, explainability, causal discovery, and machine teaching techniques to user profiling. The goal is to support decision-making by identi fying the most relevant features, clarifying causal mechanisms, and ensuring that stake holders understand why recommendations are made. Specifically, we investigate the mRMR (minimum-Redundancy-Maximum-Relevance) method for feature selection, ex amine explainability strategies such as feature importance analysis and counterfactuals, apply causal discovery to map cause-and-effect relationships, and use machine teaching to explore profile simplification. We apply this approach in four domains: (i) Marine litter: developing static profiles to identify those who could benefit from literacy interventions; (ii) Football injuries: building predictive models based on player profile dynamics to forecast risk; (iii) Energy poverty: designing models, using counterfactuals, and applying causal discovery to understand health–poverty links; and (iv) Concept complexity: using machine teaching to study profile simplification. These applications show how profiling can deliver targeted literacy interventions, prevent sports injuries, inform preventive policies in energy poverty, and improve the efficiency of user representations and concept learnability.