SILVA, S. D.; http://lattes.cnpq.br/2989521169184957; SILVA, Salatiel Dantas.
Resumo:
Points of Interest (POIs) are specific locations, such as restaurants, shopping centers, and
parks, considered relevant to users. Representing their types through computational mech anisms is crucial for developing solutions that assist in tasks such as urban planning, clus tering, and POI recommendation. Recent approaches have used high-dimensional vectors
(vector embeddings) to represent POI types based on contextual neighborhood relationships
or words associated with POIs. Such representations have overlooked the geographical fea tures present in the vicinity, such as streets, buildings, rivers, and parks. However, these fea tures can significantly contribute to a better representation of POI types. In this context, this
research proposes an approach to generate embeddings of POI types using the geographic
features present in the context of POIs. In the proposed approach, the GeoContext2Vec al gorithm has been developed and employed, which considers POI types and the geographical
features present in their context to generate a training set, preserving spatial patterns and
occurrences of the features. This set is used to train the Word2Vec and DistilBert models,
from the Natural Language Processing (NLP) area, capable of generating the embeddings
of types. As the main results obtained, it was found that the embeddings produced with
GeoContext2Vec reflect the similarity of POI types according to hierarchical structures and
people’s opinions, with matching values of approximately 98%, surpassing state-of-the-art
strategies. Furthermore, the results indicate the superiority of these embeddings in an urban
zone classification task, achieving an F-Score value of 90%. This result demonstrates that
geographical features are relevant information in the representation of POI types.