Abstract
INTRODUCTION: Interactive installation art offers immersive and participatory environments that elicit complex and multidimensional emotional experiences-encompassing sensorimotor engagement, affective resonance, and cognitive reflection. However, these emotional responses' inherently dynamic, subjective, and often pre-reflective nature poses significant challenges to their systematic prediction and computational modeling. METHODS: To address these challenges, the present study introduces an interpretable machine learning framework grounded in the Random Forest (RF) algorithm, which provides a balanced trade-off between predictive performance and model transparency, thereby aligning with the needs of theory-driven emotion research. Based on 390 valid questionnaire responses, emotional responses were operationalized along five distinct dimensions: bodily changes, sensory engagement, emotional connection, cognitive reflection, and active personalization. Predictor variables encompassed sensory stimuli, multimodal interactional features, and immersive environmental cues. Model evaluation was conducted using cross-validation and held-out test sets, applying classification and regression metrics to assess performance. RESULTS: The RF model demonstrated the highest predictive accuracy in the domains of cognitive reflection (F1 = 0.746, accuracy = 0.769) and active personalization (F1 = 0.673, accuracy = 0.705), suggesting that these cognitively mediated responses exhibit greater consistency and learnability across participants. In contrast, bodily responses proved substantially less predictable (F1 = 0.379, accuracy = 0.397), likely due to their idiosyncratic, embodied, and non-verbal nature, which may not be adequately captured by self-report measures alone. DISCUSSION: These differential results underscore the relative tractability of modeling reflective and agentic emotional states in contrast to those rooted in sensorimotor or affective processes. Moreover, the model's consistent performance across all evaluation phases affirms its suitability as an exploratory tool for investigating emotion in interactive art contexts. This study contributes to the evolving convergence of affective computing, human-computer interaction (HCI), and empirical aesthetics. The proposed framework yields actionable insights for the design of emotionally adaptive systems. Future research should consider the integration of multimodal and temporally granular data, and the ethical dimensions associated with affective adaptivity in artistic and public-facing environments.