Where You Are Is What You Do: On Inferring Offline Activities From Location Data

TL;DR

This paper demonstrates that modern machine learning algorithms can accurately infer offline activities like shopping and dining from location data, highlighting both technological capabilities and privacy concerns.

Contribution

It empirically evaluates the performance of various machine learning models in inferring offline activities from location data, revealing the effectiveness of tabular models.

Findings

F1 score > 0.9 for activity inference

Tabular models perform among the best

Highlights privacy risks of location data

Abstract

In this paper we investigate the ability of modern machine learning algorithms in inferring basic offline activities,~e.g., shopping and dining, from location data. Using anonymized data of thousands of users of a prominent location-based social network, we empirically demonstrate that not only state-of-the-art machine learning excels at the task at hand~(F1 score>0.9) but also tabular models are among the best performers. The findings we report here not only fill an existing gap in the literature, but also highlight the potential risks of such capabilities given the ubiquity of location data and the high accessibility of tabular machine learning models.