A Multigrain-Multilabel (MGML) Dataset for Smartphone-Based Human Activity Recognition

The efficacy of machine learning-based Human Activity Recognition (HAR) heavily relies on the datasets. Existing benchmark HAR datasets on smartphone accelerometer sensors provide mostly single-labeled, fine-grained activities like walking, sitting, etc. collected in lab set-up. In real life, users hardly perform an activity in isolation. Rather, the activities are often performed in sequence with non-uniform transition duration. So, for faster transitions, fine-grained data annotation is difficult and error-prone. A few existing benchmark datasets report coarse-grained activities like working, cooking, etc. that do not indicate enough information about the constituent fine-grained activities performed. HAR performance in these cases cannot satisfy real-life purposes like physical fitness prediction, or rehabilitation after surgery. To address this challenge, a Multigrain-multilabel (MGML) dataset has been designed by collecting smartphone accelerometer sensor readings from four users. Here, the coarse-grained activities are multi-labeled, and both fine-grained and coarse grained activities are covered. That means one can get information regarding the physical movements of the user from labelling information. The MGML dataset has been evaluated with four machine learning classifiers. We report a baseline classification accuracy of 95.40% for the classifiers considered. Experimentation has been conducted on the entire filtered and feature-engineered data. The effectiveness of feature engineering has also been shown.