Identifying Key Factors in Human-Computer Interaction Through Feature-Based Analysis and Sensitivity Evaluation

Abstract

For this experiment, Human-Computer Interaction (HCI) user interaction was forecasted by machine learning (ML) classification models, Random Forest Classification (RFC), and Decision Tree Classification (DTC). Two metaheuristic optimization algorithms, Fox Optimizer (FO) and Northern Goshawk Optimization (NGO), were employed to improve the precision and reliability of these predictive models. The aim was not only to build good predictive models but also to identify the most vital factors that determine the models' performance. Therefore, a sensitivity analysis was conducted using the Class Activation Mapping (CAM) method. The importance of each input variable was quantitatively assessed using Sobol's sensitivity analysis, with both the first-order effect (S1) and the total effect (ST) index being considered. The sensitivity analysis outputs showed that mouse distance, clicks per minute, and interface complexity were the most important features, with each having a normalized sensitivity index higher than 0.30. All three issues dominated the variance of the model's output, illustrating their ability to represent user interaction patterns. The optimization with FO and NGO resulted in a significant improvement in classification accuracy. Furthermore, such optimization techniques can be an effective way of improving the performance of classification models. This research is relevant to identify the most influential behavioral and interface-related factors that influence user engagement across HCI environments. Besides, the methodology employed in this study can be used as a road map for future research projects aimed at the design of user-adaptive systems and the implementation of smart interface designs based on data-driven guidelines. The results highlight the importance of using ML models optimized for sensitivity analysis in HCI prediction tasks.