Project overview
The widespread use of mobile devices has created new opportunities to study population activity and mobility at unprecedented spatial and temporal scales. Location-related information can be obtained from several sources, including Global Positioning System (GPS) traces, proximity to radio-frequency beacons, geolocated social-network data and Call Detail Records (CDRs) generated by mobile network operators.
CDRs contain metadata associated with interactions between mobile devices and the telecommunications network, such as calls, text messages and internet activity. These records include a timestamp and the radio base station handling each interaction. Once anonymised and aggregated, they provide privacy-conscious indicators of collective activity without requiring the reconstruction of individual trajectories.
The spatial resolution of CDR-derived information depends on the coverage areas of the telecommunications network. Consequently, the resulting grid cells must be understood as units of analysis rather than as measurements of the precise location of individual users. Nevertheless, the volume and temporal continuity of these data make them valuable for identifying population-level activity patterns and producing functional proxy maps of urban and regional areas.
This doctoral research investigated how supervised, multi-label and spatially aware machine-learning techniques can extract population, land-use and mobility patterns from mobile-phone and geolocation data. The work focused on interpretable methods that can operate with relatively small labelled datasets and that explicitly account for the temporal and spatial structure of the observations.
Research objectives
The project pursued the following objectives:
- Analyse existing methods for identifying population and mobility patterns using CDRs, GPS traces, radio-frequency beacons and geolocated social-network data.
- Develop supervised-learning alternatives to the unsupervised clustering techniques traditionally used to classify mobile-phone activity patterns.
- Model the mixed-use nature of urban areas through multi-label classification, allowing each spatial unit to receive several simultaneous functional and morphological descriptors.
- Improve multi-label nearest-neighbour methods by adapting the number of neighbours to the local characteristics of the feature space and the geographical context.
- Evaluate the potential of the resulting methods for urban analysis, transport planning, service provision and the study of commuting flows.
Data and methodology
The main experiments used anonymised and pre-aggregated CDR datasets released by Telecom Italia for the city of Milan and the Province of Trento. The data contain activity volumes for incoming and outgoing calls, incoming and outgoing SMS messages and internet connections, aggregated into regular spatial grids and temporal intervals.
Temporal activity signatures were constructed for each grid cell and used as input to machine-learning classifiers. Representative areas were manually labelled using satellite imagery, OpenStreetMap data and other contextual geographic sources. The considered descriptors included:
- residential and work-related functions;
- high, medium and low density urban morphology;
- forested areas in the Trento case study.
These categories were treated as area-level functional and morphological proxies. In particular, home and work labels refer to the predominant characteristics of spatial units rather than to the personal home or workplace of a particular mobile-phone user.
The methodology included supervised classification, multi-label learning, local optimisation of nearest-neighbour parameters, spatial adjacency relationships, iterative stratification, spatial block cross-validation, statistical comparison of algorithms and geographic visualisation of predictions.
Main contributions
SAMPLID: supervised classification of functional areas
SAMPLID introduced a supervised framework for classifying spatial units from aggregate CDR activity. Instead of allowing a clustering algorithm to create groups solely from similarities within the data, a representative set of manually labelled cells was used to train a classifier aligned with the intended categories.
In the Milan case study, SAMPLID was applied to distinguish residential and work-related areas. The supervised k-nearest-neighbour implementation achieved an accuracy of approximately 78.5%, outperforming the k-means and k-medoids clustering alternatives considered in the study.
Extension to heterogeneous and mountainous environments
The supervised approach was subsequently evaluated in the Province of Trento, whose Alpine terrain, dispersed settlements and extensive forested areas provide a substantially different context from metropolitan Milan.
The analysis compared support vector machines, random forests, k-nearest neighbours and neural-network models. The results demonstrated that aggregate mobile-phone activity can provide useful information not only about residential and work-related functions, but also about sparsely populated and forested environments.
MAPLID: modelling mixed-use urban space
MAPLID reformulated the problem as multi-label classification. This is particularly relevant because urban cells rarely have a single, exclusive function. A spatial unit may simultaneously contain residential buildings, workplaces and a particular level of urban density.
The multi-label formulation therefore avoids forcing each cell into a single dominant category and provides a richer representation of mixed-use urban structure. The research systematically compared problem-transformation methods, classifier chains, ensemble approaches and multi-label nearest-neighbour algorithms across the Milan and Trento datasets.
Local and spatially aware nearest-neighbour methods
The methodological research first developed ML-localkNN, which assigns an instance-specific value of (k) rather than using the same number of neighbours throughout the feature space.
This idea was subsequently extended through ML-proxkNN. The method combines two complementary notions of locality:
- feature-space locality, based on similarity between temporal mobile-phone activity patterns;
- spatial-grid locality, based on the physical adjacency of neighbouring cells.
ML-proxkNN incorporates Moore-neighbourhood relationships into the optimisation of the local (k) values. It achieved an average precision of 92.0% in Milan and 90.4% in Trento. Spatial block cross-validation was also used to assess whether performance remained robust when geographically adjacent cells were prevented from appearing in both the training and test partitions.
Cross-border commuting from geolocated social-network data
The thesis also investigated mobility patterns that cannot be captured adequately through data from a single national mobile-network operator. Geolocated social-network data were used to identify home–work mobility flows in three European cross-border regions:
- the Greater Region of Luxembourg;
- the Basque cross-border region;
- the Øresund Region.
Eighteen temporal, spatial and network-based mobility features were extracted, and CatBoost, XGBoost, random forests and k-nearest neighbours were compared. The models achieved approximately 98% movement-level accuracy on the manually validated Luxembourg sample.
A zero-shot transfer-learning experiment was then conducted by applying models trained in Luxembourg to the Basque and Øresund regions without retraining. The predicted aggregate flows were compared with available official statistics. This study should be interpreted as a methodological proof of concept for aggregate commuter-flow estimation rather than as a population-representative classification of individual commuters.
Potential applications
The resulting methods can provide complementary evidence for:
- urban and regional planning;
- identification of residential, employment and mixed-use areas;
- public-transport network design;
- analysis of commuting corridors;
- location of areas requiring additional transport services;
- tourism and cross-border mobility analysis;
- infrastructure and service planning;
- environmental, accessibility and mobility-inequality studies.
For transport planning, the inferred functional maps can help identify plausible origins, destinations and temporal patterns of activity. However, aggregate CDR indicators do not directly provide complete origin–destination demand matrices or individual travel modes. They should therefore complement rather than replace transport surveys, administrative records, ticketing data and dedicated mobility applications.
Limitations and future research
The principal limitations arise from the spatial aggregation of the telecommunications data, variations in radio-base-station coverage, uncertainty in manually assigned functional labels and the availability of representative ground truth. The produced maps are exploratory functional layers and should not be interpreted as authoritative zoning products.
Future work includes validation with data from other operators and countries, temporal transfer between different observation periods, integration with transport, land-use and environmental datasets, graph-based representations of irregular spatial units and the development of methods for estimating uncertainty in predicted functional maps.
Doctoral thesis
Author: Manuel Mendoza-Hurtado
Thesis: Identification of population patterns using advanced machine learning techniques applied to mobile phone and geolocation data
Doctoral programme: Advanced Computing, Energy and Plasmas, University of Córdoba
Supervisor: Prof. Domingo Ortiz-Boyer
Defence date: 4 June 2026
Award: Sobresaliente cum laude, with International Doctorate distinction
Selected publications
Mendoza-Hurtado, M., Järv, O., Malekzadeh, M., Karasov, O., and Ortiz-Boyer, D. (2026). “Sensing labour mobility flows of cross-border urban regions using machine learning and geolocated social network data”. EPJ Data Science, 15(1). https://doi.org/10.1140/epjds/s13688-026-00662-1
Romero-del-Castillo, J. A., Mendoza-Hurtado, M., Ortiz-Boyer, D., and García-Pedrajas, N. (2022). “Local-based (k) values for multi-label (k)-nearest neighbors rule”. Engineering Applications of Artificial Intelligence, 116, 105487. https://doi.org/10.1016/j.engappai.2022.105487
Mendoza-Hurtado, M., Romero-del-Castillo, J. A., and Ortiz-Boyer, D. (2024). “SAMPLID: A new supervised approach for meaningful place identification using Call Detail Records as an alternative to classical unsupervised clustering techniques”. ISPRS International Journal of Geo-Information, 13(8), 289. https://doi.org/10.3390/ijgi13080289
Mendoza-Hurtado, M., Cerruela-García, G., and Ortiz-Boyer, D. (2025). “A supervised approach for land use identification in Trento using mobile phone data as an alternative to unsupervised clustering techniques”. Applied Sciences, 15(4), 1753. https://doi.org/10.3390/app15041753
Mendoza-Hurtado, M., Romero-del-Castillo, J. A., and Ortiz-Boyer, D. (2026). “MAPLID: A new multi-label approach for place identification using data supplied by mobile network operators”. International Journal of Geographical Information Science, 1–24. https://doi.org/10.1080/13658816.2026.2617932