Whether a mobility event is considered anomalous depends on choices that are often left implicit. These include how we represent location, time, and activity meaning; what counts as normal; and whether we define anomalies at the level of a visit, a trip, a person, or a group. I will begin with a framework from our ongoing survey that organizes these choices across anomaly types, detectors, and benchmarks. I will share initial survey findings and use examples from our collaborative work on mobility behavior clustering, context-aware trajectory anomaly detection, and changes in individual routines to illustrate several of these categories.
The second part of the talk will examine what gives a place its meaning and how that meaning changes with spatial context. I will discuss work on learning place and region representations from language, geographic information, and the surrounding environment. I will then focus on NARA, which learns spatial context from heterogeneous vector geoentities by modeling nearby entities, their geometry, and their metric and topological relations. I will also present early results from integrating NARA with a trajectory mining model for anomaly detection and close with open questions about when context helps detect an anomaly and when it can hide one.
