Εξόρυξη δεδομένων σε μεγάλα δεδομένα κίνησης

Μανδάλης, Πέτρος; Mandalis, Petros

Data mining in big mobility data

Doctoral Thesis

Author

Μανδάλης, Πέτρος

Mandalis, Petros

Date

2026-06

Abstract

Mobility data generated by systems such as AIS, GPS, and ADS-B provide unprecedented opportunities for analyzing and forecasting the movement of vessels, vehicles, and aircraft. At the same time, these data introduce significant challenges due to their scale, irregular sampling, heterogeneity, and operational time sensitivity. This thesis addresses the broader problem of data mining in big mobility data through a coherent set of contributions spanning trajectory forecasting, aggregate flow forecasting, scalable learning, methodological deep learning, and trajectory representation. The thesis first studies mobility forecasting at the level of individual moving objects through Vessel Route Forecasting (VRF). A systematic experimental benchmark is developed in order to compare statistical, machine learning, and deep learning approaches under realistic AIS-derived conditions, with emphasis on data preprocessing, evaluation design, and comparability. The thesis then moves to aggregate forecasting through Vessel Traffic Flow Forecasting (VTFF), where traffic evolution is predicted over a spatial partition. In this setting, a structured comparison is carried out between direct approaches, which forecast aggregate flow values directly from historical traffic observations, and indirect approaches, which first forecast vessel trajectories and then aggregate them into future flows. Building on this distinction, the thesis introduces UA-VTFF, a unified forecasting framework that combines vessel-level and flow-level predictive information in order to improve robustness and predictive performance. This unified direction is extended through dUA-VTFF, a distributed and Transformer-based framework designed to address the scalability requirements of large AIS datasets and to treat computational feasibility as a core dimension of mobility forecasting rather than as an implementation detail. Beyond forecasting architectures, the thesis also contributes to methodological deep learning through KANB, a family of trainable parametric activation functions based on cubic Bézier curves. KANB is evaluated as a modular replacement for standard activation functions in forecasting networks. Finally, the thesis addresses trajectory representation and clustering in unsupervised settings. A benchmark framework is proposed for trajectory vectorization, dimensionality reduction, and clustering, enabling a systematic comparison of representation families across multiple mobility datasets. This contribution emphasizes that representation is not merely a preprocessing step, but a central factor affecting clustering quality, similarity structure, and the use of off-the-shelf scalable analytics tools for trajectory data. Taken together, the contributions of the thesis establish a unified view of mobility analytics as a problem of representation, prediction, systems design, and methodological robustness. The results show that accurate and scalable mobility forecasting requires not only effective predictive models, but also rigorous evaluation, appropriate abstraction levels, reusable deep-learning components, and representations that support both learning and large-scale analytics. Although the empirical core of the thesis is centered primarily on maritime mobility, its broader methodological perspective is relevant to mobility data mining across maritime, urban, and aviation domains.

Department

Σχολή Χρηματοοικονομικής και Στατιστικής. Τμήμα Στατιστικής και Ασφαλιστικής Επιστήμης

Number of pages

142

Language

English

URI

https://dione.lib.unipi.gr/xmlui/handle/unipi/19427

Collections

Τμήμα Στατιστικής και Ασφαλιστικής Επιστήμης

Show full item record

Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα