Inferring Activity Patterns from Sparse Step Counts Data with Recurrent Neural Networks

Abstract

As an accurate measurement of physical activity, step counts data can be collected expediently by smartphones and wearable devices. Complete and high time-resolution step counts data record the time and intensity of individuals’ physical activity in a day, and can be used to mine activity habits or to recommend customized workout plans. However, sparse step counts data are common in practice due to hardware and software limitations. Understanding the value of sparse step counts data can contribute to its application in healthcare, and also can help us design cost-effective hardware and software. In this article, we aim to infer activity patterns from sparse step counts data. We design a deep learning model based on recurrent neural networks, namely MLP-GRU, which considers bidirectional short-term dependency and long-term regularity of sparse step counts data, and implements data-driven imputation and classification. We also develop an interpretable and elastic method to obtain sparse step counts data labeled with multi-granular activity patterns to train MLP-GRU. Evaluations on real-world datasets reveal that MLP-GRU outperforms other strong baseline methods. The results also show that activity patterns can be inferred from extremely sparse step counts data with high accuracy, provided that proper granularity is used for data of different sparsity.