ANDRADE, L. H.; http://lattes.cnpq.br/0264116850662513; ANDRADE, Luiz Henrique de.
Resumen:
The large volume of data concerning trajectories enables research on different domains including trajectories of pedestrians, animals, ships, and airplanes. As a result, several machine learning applications have been proposed based on these data. To achieve the success of these applications, it is important to understand the application domain, the available techniques, and the various input features. This comprehension is fundamental for the research to evolve and obtain good results. Much has been discussed on the available techniques and their different domains; however, research on features has received little to none of the attention. Another point to be considered regarding trajectories is that in some domains the interpretability of results must be rated as most important. The need for an interpretable model creates a constraint regarding the complexity of the features used along the machine learning process. The option of automatic features generation is impracticable for different domains. This is due to the complexity of the resulting features. Thus, the feature engineering process depends more on features designed by specialists. These features are based on theories regarding their various contexts. We propose a taxonomy called TrajTax that classifies trajectory features used in the stateof-the-art. The taxonomy is intended to help researchers along the feature engineering process; this will serve as a basis for specialists to discuss the features which are currently used and to suggest new ones. For the development of the taxonomy, we conducted a survey on works based on trajectory features along different domains, and used a methodology for taxonomy construction. The features were outlined based on their definitions proposed in the state-of-the-art. The Trajtax consists of a simplified comprehension concerning the features so it can help with the model interpretability. Consequently, the main contributions of the present work encompass a survey of the features used in trajectory machine learning applications on different domains, and the creation of the TrajTax taxonomy.