AutoML Framework for Physical Activities Recognition

The rise of measuring, computing, and storage capabilities in modern information systems has led to vast amounts of data in various fields of human activities. Various algorithms for machine learning have been developed to get valuable insights from this raw data. Tuning the hyperparameters of each algorithm is necessary. Finding the right algorithm and manually tuning its hyperparameters for a given task can be very time-consuming. The methods for automating this search are within the scope of AutoML (automated machine learning). The wide use of wearable devices such as smartwatches and fitness trackers enables the gathering of various information about the physical activities of the wearer. This information, in turn, can be used to produce useful insights, such as recognizing activities being performed. AutoML can automate the process of building models for this task, but the performance of the resulting models can be very different depending on the AutoML library that performed the task. In this work, we develop an AutoML framework that unites several commonly used AutoML libraries and enables the use of common constraints and metrics. The developed framework is then used to build models for the task of human activity recognition by processing a dataset containing information collected from smart devices worn by users.