Machine learning and privacy preserving algorithms for spatial and temporal sensing
Sensing physical and social environments are ubiquitous in modern mobile phones, IoT devices, and infrastructure-based settings. Information engraved in such data, especially the time and location attributes have unprecedented potential to characterize individual and crowd behaviour, natural and technological processes. However, it is challenging to extract abstract knowledge from the data due to its massive size, sequential structure, asynchronous operation, noisy characteristics, privacy concerns, and real time analysis requirements. Therefore, the primary goal of this thesis is to propose theoretically grounded and practically useful algorithms to learn from location and time stamps in sensor data. The proposed methods are inspired by tools from geometry, topology, and statistics. They leverage structures in the temporal and spatial data by probabilistically modeling noise, exploring topological structures embedded, and utilizing statistical structure to protect personal information and simultaneously learn aggregate information. Proposed algorithms are geared towards streaming and distributed operation for efficiency. The usefulness of the methods is argued using mathematical analysis and empirical experiments on real and artificial datasets.