Novel Ensemble Models for Enhanced Accuracy in Time Series Classification: Application to Multimodal Emotion Detection

Abstract

Emotions are fundamental to the human experience and are increasingly analyzed in applications such as marketing, healthcare, and human–computer interaction. Many recent approaches to human emotion recognition rely on deep learning, which typically demands large labeled datasets and substantial computational resources and often suffers from limited interpretability. Applying classical machine-learning methods to sensor time series is more lightweight but may struggle to reach high accuracy, especially when the temporal structure is not explicitly modelled. This paper introduces three subinterval voting-based ensemble models designed for user-specific emotion classification from multimodal time-series data acquired by smartwatch inertial sensors and heart-rate measurements. Each model partitions a time window into subwindows and performs window-level voting, thereby exploiting the temporal consistency of emotional responses while remaining compatible with standard classifiers such as logistic regression and Random Forests (with or without hyperparameter tuning). The models are evaluated on a public smartwatch emotion benchmark dataset under both binary (happy vs. sad) and three-class (happy, sad, neutral) settings. The relative accuracy improvement over the corresponding baseline reported in prior work ranges from 4.68% to 26.05%, with a mean gain of 12.34%. For the three-class tasks, improvements range from 11.17% to 37.10%, with a mean gain of 21.63%. Within the evaluated experimental setting, these results show that the proposed subinterval ensembles consistently enhance performance while remaining model-agnostic and compatible with standard user-specific classification pipelines in sensor-based emotion recognition.