Naïve Adaptive Probabilistic Sensor Fusion for Enhancing Context Recognition

Wearable and non-wearable smart devices are capable of capturing huge amount of information from their users, this is possible because of a variety of sensors installed on them. Sensor fusion technology can enable us to obtain a holistic picture of the users' context and accurately monitor their state, although it is very challenging to classify human context in real world. Handling multiple classes, uncertainty related to machine learning models, class-imbalance and large feature space are still some issues which need to be resolved, however recent researches propose different probabilistic models. Towards addressing these issues, in this study we explore a new probabilistic model fusion approach called Naive Adaptive Probabilistic Sensor (NAPS) fusion. This sensor fusion technique is capable of addressing uncertainty for multi- class classification in machine learning problems with large imbalanced Human Activity Recognition (HAR) dataset. Our empirical evaluation acclaims that NAPS fusion outperforms conventional sensor fusion technologies and enhances context recognition in natural environment. Our approach avoided dimensionality reduction techniques by developing structured feature-sets, which helps to resolve the class-imbalance issue. For multi-class classification task on Extrasensory dataset, NAPS fusion outperforms different conventional sensor fusion models by 5%-56% in f1 score and 3%-9% in balanced accuracy.