A Novel Method for Wearable Activity Recognition with Feature Evolvable Streams

Wearable activity recognition plays an important role in human health monitoring. Traditional wearable activity recognition models are trained in an offline mode with static and pre-defined sensor configurations. However, in real scenarios, data arrives in streams and wearable sensors dynamically appear or disappear, resulting in corresponding changes in the feature space, which is referred to as feature evolution. Addressing the issue of feature evolution is a significant challenge in wearable activity recognition. In this paper, we propose a new method, namely Online Learning method for Feature Evolvable Streams (OLFES). OLFES learns the optimal model depth online according to the complexity of the data stream, recovers the old features through the feature space generation strategy, and finally integrates the prediction results according to a weighted combination strategy. Extensive experimental results on data science datasets and activity recognition datasets demonstrate the feasibility and effectiveness of the proposed method.