Attention-Based Deep Learning Framework for Human Activity Recognition With User Adaptation

Sensor-based human activity recognition (HAR) requires to predict the action of a person based on sensor-generated time series data. HAR has attracted major interest in past few years, thanks large number applications enabled by modern ubiquitous computing devices. While several techniques hand-crafted feature engineering have been proposed, current state-of-the-art is represented deep learning architectures that automatically obtain high level representations and use recurrent neural networks (RNNs) extract temporal dependencies input. RNNs limitations, particular dealing with long-term dependencies. We propose novel framework, \algname, purely attention-based mechanism, overcomes limitations state-of-the-art. show our proposed architecture considerably more powerful than previous approaches, an average increment, $7\%$ F1 score over best performing model. Furthermore, we consider problem personalizing models, which great importance applications. simple effective transfer-learning strategy adapt model specific user, providing increment $6\%$ predictions for user. Our extensive experimental evaluation proves significantly superior capabilities framework effectiveness user adaptation technique.