LSTM-XGB: A New Deep Learning Model for Human Activity Recognition based on LSTM and XGBoost

Recognition of human activity focuses on identifying various human motions and behaviors using data acquired from multiple types of sensors, a branch of computer science. Considering time series data, deep learning algorithms show promise compared to laborious handcrafted feature extraction methods, which are highly dependent on the quality of the used domain parameters. This paper proposes a novel deep learning model for HAR issues called the Long Short Term Memory Neural network with eXtreme Gradient Boosting (LSTM-XGB), developed on recurrent neural networks with eXtreme Gradient Boosting (XG-Boost). The LSTM-XGB comprises LSTM layers to understand the characteristics of the input and is capable of learning features automatically, followed by XGBoost in the final layer to forecast the class labels, which serves to enhance detection capability. When sufficient circumstances are satisfied, the LSTM-XGB model can be simpler by lowering the number of parameters. It is not essential to readjust the weighting factors throughout a backpropagation phase. Investigations using smartphone sensor data were carried out using a publicly available standardized HAR dataset called the UniMiB-SHAR dataset, which contains activities associated with everyday life and activities related to falls. The experiments show that the LSTM-XGB provides much superior identification capabilities than other standard deep learning techniques, with the highest accuracy of 92.59%. Furthermore, the LSTM-XGB models surpass the existing best-practice models on the same dataset.