A Comparative Study of Classical Machine Learning and Deep Learning Approaches for Human Behavior Detection Using Multisensor Data

Human behavior detection using multisensor data is increasingly important in healthcare, assistive technologies, and smart environments. This study provides a systematic comparative analysis of classical machine learning (ML) and deep learning (DL) approaches for behavior recognition using the CMI (Cognitive Monitoring Initiative) dataset, which includes multivariate time-series signals from wearable sensors and demographic metadata. Four experimental pipelines were designed and evaluated, incorporating classical models (CatBoost, XGBoost, LightGBM, ExtraTrees) and DL architectures (Bidirectional LSTM, Transformer). A central novelty of this work is the integration of statistical and frequency-domain feature engineering with a multi-stage feature selection pipeline—mutual information, embedded Random Forest filtering, recursive feature elimination (RFE), and principal component analysis (PCA). This process significantly improved the performance of classical models by reducing noise and enhancing generalization. CatBoost, the most effective model, reached a macroF1 score of 88.00%. LightGBM and XGBoost also performed competitively under cross-validation. Deep learning algorithms battled data scarcity and class imbalance on the other hand, resulting in exceedingly low macroF1 ratings (~0.81%). These results highlight how critical feature curation and thorough validation are in modest HAR projects. Future studies should investigate hybrid ML–DL ensembles, sophisticated data augmentation, and lightweight DL models designed for wearable and IoT devices' deployment.