Applications of Machine Learning to Location Data

Abstract

Positioning devices are generating location data at an unprecedented pace. Coupled with the right software, these data may enable a virtually unlimited number of valuable services. However, to build such software, there is a need for sophisticated algorithms that can extract the relevant information from location data. In this thesis, we use machine learning to develop such algorithms for three fundamental location-based problems.

First, we introduce a new graphical model for tracking radio-tagged animals and learning their movement patterns. The model provides a principled way to combine radio telemetry data with an arbitrary set of spatial features. We apply our model to real datasets and show that it outperforms the most popular radio telemetry software package used in ecology, produces accurate location estimates, and yields an interpretable model of animal movement.

Second, we develop a novel collaborative ranking framework called Collaborative Local Ranking (CLR), which is designed to solve a ranking problem that occurs frequently in the real-world but has not received enough attention in the scientific community. In this setting, users provide their affinity for items via local preferences among a subset of items instead of global preferences across all items. We justify CLR with a bound on its generalization error and derive an alternating minimization algorithm with runtime guarantees. We apply CLR to a venue recommendation task and demonstrate it outperforms state-of-the-art collaborative ranking methods on real datasets.

Third, we design two Bayesian probabilistic graphical models that predict users' future geographic coordinates based on sparse observations of their past geographic coordinates. Our models intelligently share information across users to infer their locations at any future weekhour, determine the number of significant places and the spatial characteristics of these places, and compute the conditional distributions that describe how users spend their time at these places. We apply our models to real location datasets and demonstrate that, despite the sparsity, they provide accurate representations of users' places and outperform existing methods in estimating users' future locations.