Recommendations for the creation of datasets in support of data driven activity recognition models

In the last decades, datasets have emerged as an essential component in the process of generating automated Activity Recognition (AR) solutions. Nevertheless, some challenges still remain: the lack of recommendations about which kind of information should be represented inside a dataset has resulted in the implementation of a variety of different non-standardized formalisms. On the other hand, this information is usually not sufficient to fully characterize the dataset. To address these challenges, this paper introduces a series of recommendations in the form of a dataset model with a well-defined semantic definition, for supporting those who are responsible for the creation, documentation and management of datasets. In addition, in order to better characterize datasets from a statistical point-of-view, we describe eight statistical analyses which should be included as additional measures within the dataset itself. We have validated our concepts through retrospectively analyzing a well-known dataset.