Using data augmentation, active learning, and visual analytics for learning with limited examples on mobility data sets

点击翻译按钮获取中文摘要

The increasing access to positioning devices technologies such as smartphones and GPS-enabled cameras has resulted in vast volumes of mobility data collected, stored, and available for analysis. However, it is difficult to find data sets containing annotated trajectories since labels tend to indicate a specific behavior whose identification depends on human interpretation. Labels are essential in supervised learning problems to predict an outcome, such as the transportation mean used by a user carrying a smartphone or if a vessel is performing a fishing activity. Understanding the moving object behavior depends on multiple factors and, often, cannot be automatically inferred. Therefore, the annotation process may be complex and time-consuming even for domain experts since trajectory data has a vast volume and, in many cases, several attributes (high dimensionality). Consequently, reducing the labeling effort is of interest to researchers and industries working in marine transportation, tourism, wildlife monitoring, and traffic management. A promising approach to address the lack of labels is to use data augmentation techniques that create new data samples from known patterns in order to expand the decision boundary of a classification model. Another approach is to use active learning strategies that help to select a subset of examples from an unlabeled pool of data so that obtaining annotations will result in a maximal increase of model performance. Current strategies cannot be used directly with trajectory data due to their spatiotemporal dependencies. Trajectory data tend to be large and high dimensional, making it difficult to annotate it from the user's perspective. Assembling interactive tools for annotating trajectories to effectively assist the user in coping with all the aforementioned aspects during the labeling process can be done using visual analytics. Visual analytics uses interactive graphical interfaces to help make decisions more efficiently and effectively by merging interactivity with automated visual analysis. Therefore, this research program aims to develop new data augmentation, active learning, and visual analytics strategies and tools to allow the mobility data mining field to learn robust models with a limited set of examples. Learning with a limited set of examples is one of the most challenging setups in machine learning. This program will train several undergraduate and graduate students who will benefit from realistic and industry-relevant thesis' topics in the mobility data analysis domain. Government departments and Canadian industries working with these topics will benefit from the outcomes of this program (Highly Qualified Personnel (HQP), production of software systems, and libraries). Moreover, this program will provide new solutions to current applied science and technology challenges involving mobility data, decrease costs involved in creating labeled mobility datasets, and provide more accurate services.