Mitigating Interpersonal Variability in Machine Learning Models for Human Behavior Monitoring

Abstract

Machine learning models for human behavior monitoring typically rely on population-level data, assuming that physiological patterns generalize well across individuals. However, this assumption often fails due to interpersonal variability - differences in individuals' physiological responses and subjective interpretations influenced by personal, cultural, and psychological factors. This research first investigates how such variability impacts model performance in two scenarios: sleep monitoring, where label variability arises because individuals assign different subjective ratings to similar physiological patterns, and transport mode recognition, where signal variability occurs as individuals exhibit distinct sensor signals for the same activity. Our empirical findings show that population models struggle to generalize, especially in tasks involving subjective assessments. To overcome these challenges, we firstly propose generative modeling techniques to augment limited user-specific data, and secondly introduce a clustering-based strategy that enables scalable personalization by grouping users according to both physiological signal patterns and labeling behavior.