Effective Trajectory Imputation using Simple Probabilistic Language Models

Abstract

Trajectory data analysis refers to the systematic exploration of spatial and temporal movement patterns in trajectory datasets. Missing trajectory points pose a challenge as they affect downstream tasks that rely on these datasets, such as public transportation management, wildlife monitoring, and urban planning. Diverse approaches, from statistical methods to deep learning models, have been proposed to address the issue of missing data points in trajectories. In this thesis, we explore two approaches based on relatively simple probabilistic language models in order to address this problem, also known as trajectory imputation. Using a grid-based representation, we assign tokens to each point in a trajectory, representing each trajectory as a sequence of tokens akin to a sentence in natural language. This allows the application of language models for predicting missing points (tokens). Our experiments using a real dataset of over 200,000 taxi trips show that we can fill gaps of up to 2km with 85% precision. Furthermore, compared to large language models, probabilistic language models for imputing trajectory points offer a much simpler technique and enhance the interpretability of the results.