mobiml. This demo paper introduces MobiML, a new library that aims to help scientists and engineers with developing mobility ML solutions using trajectory data. We also demonstrate how MobiML can speed up ML development workflows using the example of a reproduction of the workflow for training a GeoTrackNet trajectory anomaly detection model. MobiML is available at: https://github.com/movingpandas/ ∗Corresponding author.
The Mobility Data Science ecosystem is incredibly diverse
with numerous heterogeneous data sources and use cases.
Common machine learning (ML) use cases include [
A major hurdle in Mobility Data Science “is that existing
ML and analytics tools [...] do not support location and
mobility as base data types to reason about. [...] This raises
a fundamental and big question on what are the analysis
primitives and common building blocks for applications
that could shape a framework of ML-based mobility data
analysis?” [
To fill this gap, we present MobiML, a framework for learning from movement data that proposes essential building blocks to build ML solutions based on movement trajectory data. The following section provides information on related libraries. Then we present MobiML and its components, before we present usage examples, and finally present perspectives on future developments.
To the best of our knowledge, existing ML libraries for spatiotemporal data focus on remote sensing imagery and similar gridded datasets:
GeoTorchAI [
TorchGeo [
In the development of MobiML, we have taken inspiration from both GeoTorchAI and TorchGeo and adapted the concepts to trajectory data.
Published in the Proceedings of the Workshops of the EDBT/ICDT 2025 (M. Dragaschnig) (M. Dragaschnig)
CEUR
ceur-ws.org
GeoPandas [
The goal of MobiML is to enable the eficient development of ML solutions by providing mobility-aware building blocks for ML workflows. This means that the MobiML tools are aware of the spatiotemporal nature of trajectory data and automatically account for them. MobiML includes the following components: datasets, preprocessing tools, samplers, transformation tools, and models, which are shown in Figure 1 and described in more detail in the following.
This module contains classes for handling popular movement datasets. These classes serve to facilitate model development by providing straightforward access to common public datasets and also serve as templates for custom datasets classes that developers may want to create. Dataset classes provide a standardized interface including functions to access the data in the form of Pandas DataFrames, ryCollections, as well as to create static and interactive plots of the data.
Every class provides information on where to get access to the respective dataset (where to download it). The loading function then takes care of the multitude of diferent input formats, including pickle, csv, feather, zipped csv files, or geo data formats supported by GeoPandas such as Shapeifle, GeoPackage, or GeoJSON and maps the spatiotemporal information and mover IDs to the standardized structure.
So far, we have implemented six dataset classes based on real-world public mobility datasets from the:
and France BrestAIS, • Maritime domain: AIS data from Denmark AISDK • Urban domain: tracks of cyclists CopenhagenCyclists, buses DelhiAirPollution, and taxis PortoTaxis, and • Ecology domain:
MovebankGulls.
tracks of migratory birds
This module contains tools to preprocess movement data to generate datasets that are ready for ML development. • TrajectoryDownsampler to reduce the number of points in a trajectory to a certain target sampling interval. • TrajectoryEnricher to add features such as speed, direction, and acceleration in a variety of supported units. • TrajectoryFilter to remove trajectories based on a their number of points or their speed. • TrajectorySplitter to split long trajectories into shorter subtrajectories, for example, based on observation gaps • Normalizer to min-max normalize latitude, longitude, speed, and direction values in dataset. • StationaryClientExtractor to extract subsets of the data based on static locations (provided as GeoDataFrame). • MobileClientExtractor to extract subsets of the data based on moving locations (provided as MobiML Dataset) using PyMEOS.
This module contains tools for sampling movement data
while accounting for its spatiotemporal characteristics. The
developed sampling tools include:
• MoverSplitter to split a dataset ensuring that
trajectories of a given percentage of the movers are
assigned to the test set. The remaining mover
trajectories are assigned to the train set.
• RandomTrajSampler to randomly sample
trajectories, targeting an equal spatial distribution based on
user-defined grid (i.e. equal number of trajectories
per cell, based on start points), inspired by [
This module contains various transformation operations
that can be applied to datasets. Transforms convert a
mobiml.Dataset into a diferent data structure that conforms to
model-specific requirements. The developed transformation
tools include:
• DeltaDatasetCreator to convert absolute
locations and timestamps into relative changes, based
on [
This module contains example models used to demonstrate
mobility ML workflows. Included are:
• GeoTrackNet: Anomaly detection in maritime trafic
patterns, as presented in [
The use of these components is documented through a series of example notebooks that demonstrate individual components as well as ML workflows from data preprocessing to model training and inferencing.
In this section, we present how MobiML can be used to train
GeoTrackNet using Danish AIS data1. The original
GeoTrackNet paper [
Using MobiML’s AISDK dataset class, we can easily load one of the CSV files provided by the Danish Maritime Authority, as shown in Figure 2.
To preprocess the data, we leverage classic Pandas DataFrame operations and MobiML tools as follows: 1. Filter the dataset to our desired vessel types by directly working with the AISDK object’s DataFrame (Figure 3), 2. Split trajectories at observation gaps using
TrajectorySplitter (Figure 4) since these gaps would mess up what the model learns, 1for the full notebook see https://github.com/movingpandas/mobiml/ blob/main/examples/mobiml-geotracknet.ipynb 3. Drop trajectories that have too few points using
TrajectoryFilter (Figure 5), and finally 4. Reduce the dataset size using
TrajectoryDownsampler (Figure 6).
Since every preprocessing tool takes a Dataset as input and produces a Dataset as output, they can be chained in a modular way to support diferent needs and workflows.
To prepare the training stage, we split the dataset using the TemporalSplitter (Figure 7). In this example, we use the default: 70/20/10 split. The split subsets are then transformed into a format suitable for training the model. This training itself is controlled by a dedicated configuration, as shown in Figure 8.
Once the model is trained, the training result can be visualized on a map, as shown in Figure 9.
Finally, using the trained model, we can perform the inference by running the GeoTrackNet contrario detection step which flags anomalous trajectories such as the example shown in red in Figure 10. In this paper, we introduced MobiML, a new library designed to facilitate the development of machine learning (ML) solutions in the Mobility Data Science domain. Our demonstration of the GeoTrackNet anomaly detection model highlights MobiML’s ability to support reproducibility and eficiency in mobility-focused ML research. By providing modular components for preprocessing, transformation, sampling, and modeling, MobiML simplifies the development of ML worklfows, reducing the efort required to implement key steps. This should also help improve the reliability of research code, since we have ensured that MobiML is covered by unit tests, reinforcing its usability for the broader research community.
Moving forward, future developments should focus on expanding functionality and improving interoperability. Enhancements may include new dataset types, such as flowbased mobility data, as well as improvements in tool harmonization and packaging via PyPI and Conda-Forge. Additionally, we invite ML developers in the Mobility Data Science community to engage with us, contribute to MobiML, and provide feedback to refine its architecture. One of the major challenges in developing MobiML has been managing incompatible Python environments, given the specific requirements of various ML models and libraries. This issue, which afects reproducibility across the ML field, calls for collaborative eforts to find sustainable solutions. Addressing this challenge will be essential to ensuring scalability and long-term usability of MobiML in diverse research and application settings.
This work is mainly funded by the EU’s Horizon Europe research and innovation program under Grant No. 101070279 MobiSpaces.
During the preparation of this work, the authors used ChatGPT in order to paraphrase and reword. After using this tool/service, the authors reviewed and edited the content as needed and takes full responsibility for the publication’s content.