• classification
• interpretable
• ml
• trajectory
• xai

With the increase in storage device capacities and data transmission, it becomes easier to collect data. Among the different data types collected, a large space is reserved for mobility data as it is used in various applications ranging from location-based to route planning systems.
The most accurate and effective algorithms these systems use rely on complex statistical models humans cannot understand.
Researchers have recently proposed various approaches to mitigate the interpretability problem in recent years.
Indeed, these approaches formed a new branch of Artificial Intelligence called eXplainable AI (XAI).

Unfortunately, although researchers, business companies, and governments exploit mobility data to make decisions that impact lives in many ways, mobility data analysis is an under-experienced topic. There are few interpretable models and even fewer XAI approaches for black-box models at the current stage.
In this thesis, we propose three machine learning algorithms for GPS trajectory classification.
Our idea is to make some effective and intuitive transformations to the input data to use a simple interpretable model without losing performance.
Interpretability in our methods comes from the ability to highlight the parts of a trajectory that made the greatest contribution.
All proposed models attempt to extract discriminating subtrajectories using techniques from pattern mining, clustering and time series fields and apply transformations to the mobility data similar to those used for time series classification.
We also evaluate our proposals using real-world datasets to assess how well the different ideas perform with this particular data type.
The results show that one of our algorithm is an order of magnitude faster than the other proposals while maintaining comparable if not better accuracy than non-interpretable state-of-the-art classifiers.