Feature Encoding by Location-Enhanced Word2Vec Embedding for Human Activity Recognition in Smart Homes

Human Activity Recognition (HAR) in Smart Homes (SH) is the basis of providing automatic and comfortable living experience for occupants, especially for the elderly. Vision-based approaches could violate occupants' privacy and wearable sensors based approaches could be intrusive with their daily activities. In this study, we proposed an NLP-based feature encoding for HAR in smart homes by using the Word2Vec word embedding model and incorporating location information of occupants. We used the NLP approach to generate semantic and automatic features directly from the raw data that significantly reduced the workload of feature encoding. The results showed that both Word2Vec embedding and location-enhanced sequences can significantly improve the classification performance. Our best model which used both Word2Vec embedding and location-enhanced sequences achieved an accuracy of 81% and a weighted average F1 score of 77% on the test data with Sensor Event Windows (SEW) size of 25. This size is considered as a small SEW size which can be applied better to real-time classification due to the short latency.