Modelo Computacional Cognitivo para la predicción de preferencias de usuario, basado en análisis de emociones y preclasificación de contenido,orientado a la visualización de datos

Abstract

The rapid expansion of the digital ecosystem has intensified people’s exposure to massive volumes of content, complicating the timely selection of what is truly relevant. This thesis addresses that challenge through affective personalization: it proposes and validates a cognitive computational model able to anticipate user preferences from emotion analysis measured with electroencephalography (EEG) and integrated with standardized selfreport (SAM). The approach articulates three components: (i) reproducible preclassification of visual stimuli via an MLOps pipeline,(ii) a formalized and time synchronized EEG→SAM transformation heuristic, and (iii) personalized prediction with Gaussian Process Regression (GPR). All of this runs in the EPRA web prototype, backed by APIs and a traceable experimental database. Empirical validation was conducted using a case study of information-rich images with high affective load related to the Colombian armed conflict, presented in a controlled web environment. A total of 15 students (16–25 years) completed 5 sessions each, with 5 stimuli per session (375 observations in total). The timing per stimulus (5s baseline, 30s exposure, 15s SAM) guaranteed deterministic alignment between EEG and self-report. The results show positive concordance between measurement routes: global EEG–SAM correlation of r ≈ 0,53 for valence and r ≈0,36 for arousal, with mean absolute discrepancies MAE≈1,6 (1–9 scale). We observed inter-subject heterogeneity consistent with individualized affective profiles, supporting user-specific GPR models. In addition, there was evidence of early adaptation (improvement across the first sessions), while maintaining end-to end data traceability and reproducibility. The thesis contributions are fourfold: (i) a conceptual integration of neurophysiological (EEG) and subjective (SAM) routes within a preference oriented model; (ii) methodological formalization of the EEG→SAM heuristic, bias diagnostics, and longitudinal evaluation with robust metrics; (iii) technical materialization in EPRA (MLOps, Clean Architecture, component-based frontend, APIs, and data schema); and (iv) empirical evidence with a documented, replicable dataset. Taken together, the model and prototype demonstrate feasibility and utility for personalizing information visualization based on emotional state, and provide a transferable foundation for domains such as visual analytics, education, and neuromarketing.