Optimizing LSTM with Grid Search and Regularization Techniques to Enhance Accuracy in Human Activity Recognition

Abstract

This study aims to enhance the accuracy of Long Short-Term Memory (LSTM) models for human activity recognition using the UCI Human Activity Recognition (HAR) dataset. The dataset comprises time-series data from accelerometer and gyroscope sensors on smartphones worn by 30 volunteers as they performed everyday activities such as walking, climbing stairs, descending stairs, sitting, standing, and lying down. Optimization was carried out using Grid Search for hyperparameter tuning and L2 regularization to prevent overfitting. The results show that the optimized LSTM model improved accuracy from 92.33% to 94.50%, precision from 93.12% to 94.61%, recall from 92.33% to 94.50%, and F1-score from 92.32% to 94.51% compared to the standard LSTM model. Despite these improvements, the study encountered several challenges, particularly in tuning hyperparameters, which required significant computational resources and time due to the complexity of the search space. Additionally, balancing regularization to prevent both underfitting and overfitting proved to be a delicate process. Further limitations include the model's performance variability with different sensor placements and potential overfitting to specific activity patterns. However, the implementation of hyperparameter optimization and regularization proved effective in improving the model's ability to recognize human activity patterns from complex sensor data. Therefore, this approach holds significant potential for broader applications in sensor-based human activity recognition systems, though further research is needed to address these limitations and generalize the findings.