Enhancing user identification through batch averaging of independent window subsequences using smartphone and wearable data

Abstract

Throughout daily life, individuals partake in various activities such as walking, sitting, and drinking, often in a random manner. These physical activities generally exhibit similar patterns across different people, posing a challenge for identifying users using smartphone and wearable data. To tackle this issue, we have developed a new model called Batch Averaging Probabilities (BAP). Our approach involves segmenting input sequences into separate windows, independently classifying each segment, and then averaging the probabilistic predictions to make the final decision. The BAP method introduces the concept of primary patterns, which are the smallest meaningful sequences. It effectively deals with the random order of primary patterns within mixed patterns. Our work includes theoretical evidence supporting the BAP method, showcasing its ability to minimize prediction variance and enhance model accuracy. Additionally, the model’s training algorithm employs a unique approach. Model selection and regularization are based on the averaged loss of segments, reducing overfitting and improving performance without the complexity associated with using an ensemble of neural network models. We evaluated the effectiveness of our proposed method using accelerometer and gyroscope data from diverse user activity datasets including UIFW, WISM, HOP, CLD, RSSI, DI, DB2 and HAR, demonstrating significant performance improvements over state-of-the-art models. Specifically, our approach outperforms DB2 by 1.08%, HAR by 7.67%, and DI by 14.76% in terms of accuracy.