News

Location data is extremely sensitive, revealing where we spend how much time, what we like to do in our free time, and how we spend our money. Our new study titled “Where you go is who you are – A study on machine learning based semantic privacy attacks” sets out to answer a critical question: Just how much “semantic information” – that is, detailed and potentially sensitive insights into personal behavior – can be extracted from someone’s location data using machine learning techniques, especially when this data might be inaccurate or incomplete?

To address this question, we conducted a study using a dataset of location visits from a social network, Foursquare, where the category of the location (e.g., “dining” / “sports” / “nightlife”) is known. We simulate the following scenario: An attacker only knows the geographic coordinates and time of a location visit, and trains a machine learning model to infer the place category. Crucially, the geographic coordiates can be imprecise due to GPS inaccuracies or protection measures. This is simulated by obfuscating the true geographic coordinates with varying radius.

Our experiments reveal a significant privacy risk even with imprecise location data. Although the accuracy of the attacker decreases exponentially with the obfuscation radius, there remains a high privacy loss even if the coordinates are perturbed by up to 200m. A certain risk remains from the temporal information alone. These findings highlight the privacy risks associated with the growing databases of tracking and spatial context data, urging policy-makers towards stricter regulations.

Check out our paper and source code!

Our new paper entitled “Evaluating geospatial context information for travel mode detection” was accepted at Journal of Transport Geography and is now available (open-access!) online.

How much does geospatial context information contribute to travel mode detection?

Our latest study reveals that geospatial network features, such as distance to the road network, are more critical than motion features, such as speed and acceleration, when classifying an extensive list of travel modes. Still, most land-use and land-cover features barely contribute to the task. The results are based on our extensive context representation reviews and the proposed analytical pipeline to assess the contribution of geospatial context information based on a random forest model and the SHapley Additive exPlanation (SHAP) method.

The study provides valuable guidance for feature selection, effective feature design, and building efficient travel mode detection models.

Check out the paper online and the corresponding code on Github!

Our new paper entitled “Context-aware multi-head self-attentional neural network model for next location prediction” was accepted at Transportation Research Part C: Emerging Technologies and is now available (open-access!) online.

We present a multi-head self-attentional (MHSA) neural network that integrates location features, temporal features, and functional land use contexts for next location prediction. This comprehensive approach effectively captures movement-related spatio-temporal information, leading to state-of-the-art performance on GNSS mobility datasets.

Our analysis demonstrates that training the model on population data yields superior results by learning from collective movement patterns, surpassing the capabilities of individual-level models. Moreover, we emphasize the significance of recent past movements and weekly periodicity, showing that learning from a subset of historical mobility is sufficient to obtain an accurate location prediction result.

The proposed model represents a pivotal advancement in accurate and interpretable individual mobility prediction, and can be readily applied in downstream applications, including planning on-demand transport services, implementing mobility incentives, and suggesting alternative mobility options.

Check out the paper online and the corresponding code on Github!