Combining Public Machine Learning Models by Using Word Embedding for Human Activity Recognition
Koichi Shimoda, Akihito Taya, Yoshito Tobe · 2021
Many types of human activity recognition (HAR) techniques have been developed based on machine learning (ML) using diverse sensors in our daily life, such as cameras, microphones, and acceleration sensors on smartphones and wearable devices. For this background, there are many pre-trained ML models of HAR available on the Internet. In addition, there are cases where multiple ML models are utilized in combination for HAR. Although combining multiple ML models can be expected to improve classification accuracy, it requires additional training to combine them, such as stacking. This paper proposes a method to integrate published and trained HAR ML models without additional training by utilizing word embedding, e.g., Glove and Word2vec. Word embedding enables to obtain plausible results by arithmetic operations on the labels output by multiple ML models. Utilizing the proposed integration method, a HAR system is developed, which implements publicly-available ML models. Evaluations using Glove are performed, and their results confirm the proposed method is able to improve classification accuracy compared to using each ML model alone.