# Benchmarking Classical, Deep, and Generative Models for Human Activity Recognition

**Authors:** Md Meem Hossain, The Anh Han, Safina Showkat Ara, Zia Ush Shamszaman  
**arXiv:** [2501.08471](https://arxiv.org/abs/2501.08471)  
**Date:** 2025-01-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.08471v1/

## Key points

- Human Activity Recognition (HAR) is important due to sensor-equipped devices and large datasets
- Evaluation of three model categories: classical machine learning, deep learning architectures, and Restricted Boltzmann Machines (RBMs)
- Assessment of various models including Decision Trees, Random Forests, CNNs, and DBNs using key benchmark datasets
- CNN models demonstrate superior performance across datasets, especially on the Berkeley MHAD dataset
- RBM-based models show promise for feature learning
- Traditional machine learning models like Random Forest perform well on smaller datasets but struggle with larger and complex data
- Deep learning approaches excel in understanding complex patterns over time
- Paper provides a detailed comparison to help researchers choose the most suitable model for HAR tasks

## Summary

Human Activity Recognition (HAR) has become increasingly important due to the widespread use of sensor-equipped devices and large datasets. This paper evaluates the performance of three categories of models: classical machine learning, deep learning architectures, and Restricted Boltzmann Machines (RBMs) using five key benchmark datasets of HAR (UCI-HAR, OPPORTUNITY, PAMAP2, WISDM, and Berkeley MHAD). Various models including Decision Trees, Random Forests, Convolutional Neural Networks (CNN), and Deep Belief Networks (DBNs) are assessed using metrics such as accuracy, precision, recall, and F1-score for a comprehensive comparison. The results indicate that CNN models demonstrate superior performance across all datasets, particularly excelling on the Berkeley MHAD dataset. Classical models like Random Forest perform well on smaller datasets but face challenges with larger and more complex data. RBM-based models show promise for feature learning. This paper offers a detailed comparison to assist researchers in selecting the most suitable model for HAR tasks. The subsequent sections delve into the state-of-the-art in HAR research starting from traditional machine learning techniques to the evolution of deep learning models and recent advancements with generative models. The methodology section details data collection and preparation processes as well as model selection and configuration. Performance evaluations of each model on multiple datasets are presented in detail. Additionally, prior studies related to activity recognition systems are reviewed to provide context on the development of HAR methodologies. Early works focused on classification and recognition using classical machine learning methods while recent studies have explored various machine learning models including Decision Trees, Random Forests, Gradient Boosting DTs, Logistic regression Linear SVCs,and RBF SVM classifiers for identifying human activities in daily life. Furthermore,d eep learning methodologies have been applied for feature extraction in conjunction with traditional machine learning classifiers to enhance activity recognition accuracy. Traditional machine learning algorithms excel in scenarios where interpretability is crucial but may struggle with understanding complex patterns over time. In contrast, deep learning approaches have shown promise in automatically learning from raw data and comprehending intricate patterns that evolve over time. In conclusion, the performance evaluation of various models on multiple benchmark datasets highlights the superiority of CNN models for HAR tasks. However, RBM-based models show potential for feature learning and traditional machine learning algorithms still have their place in scenarios where interpretability is key. The paper provides a comprehensive comparison to aid researchers in selecting the most suitable model for their specific HAR task.