Citas bibligráficas
Galarreta, A., (2025). Spatiotemporal predictions from computational data substructures [Pontificia Universidad Católica del Perú]. http://hdl.handle.net/20.500.12404/29640
Galarreta, A., Spatiotemporal predictions from computational data substructures []. PE: Pontificia Universidad Católica del Perú; 2025. http://hdl.handle.net/20.500.12404/29640
@phdthesis{renati/1659341,
title = "Spatiotemporal predictions from computational data substructures",
author = "Galarreta Asian, Ana Paula",
publisher = "Pontificia Universidad Católica del Perú",
year = "2025"
}
This research explores diverse strategies for predicting users’ future locations and transactions, considering both spatial and temporal information. The study employs pattern mining and deep learning techniques to enhance predictive capabilities. Pattern mining is utilized to forecast upcoming transactions, introducing the innovation of augmenting each item with a location tag and organizing items within specific time windows. Also, a combination of deep learning with graph structures is used to predict future locations, employing the Transformer architecture to reduce training times. A significant contribution is the introduction of the Spatiotemporal Patricia Trie (SPT), offering a novel data structure for representing temporal and spatial frequent patterns. The Normalized Spatial Itemset Similarity (NSIS) metric, which evaluates predictions by considering both specific events and their associated locations, is also presented. Experiments on a private banking dataset demonstrate SPT’s potential for generating meaningful predictions about next purchase locations and categories. In addition, this work presents the Attention and Possible directions for TRAJectory prediction (APTraj) model, which leverages the Transformer’s self-attention mechanism and a directed graph of the road network to predict users’ future whereabouts. An improvement of the AP-Traj algorithm, called AP-Traj2, is also presented. AP-Traj2 emphasizes graph construction and sequence preprocessing, resulting in reduced training times and improved prediction accuracy across five distinct location datasets. This allows generalization to location data not previously represented by a graph, extending the applicability of the model. Overall, the proposed methods represent significant advances in predictive modeling for user activity and location prediction.
IMPORTANTE
La información contenida en este registro es de entera responsabilidad de la universidad, institución o escuela de educación superior que administra el repositorio académico digital donde se encuentra el trabajo de investigación y/o proyecto, los cuales son conducentes a optar títulos profesionales y grados académicos. SUNEDU no se hace responsable por los contenidos accesibles a través del Registro Nacional de Trabajos de Investigación – RENATI.