Defensa de maestría de Facundo Lezama 23/11

Defensa de maestría de Facundo Lezama 23/11

de Claudina Rattaro -
Número de respuestas: 0


Es un gusto para mí invitarlos a la defensa de maestría en Ingeniería 
Eléctrica de Facundo Lezama, a quien tuve el placer de dirigir junto con 
Germán Capdehourat. Se trata de una continuación de la línea de 
localización indoor que llevamos hace unos años junto con Germán, esta 
vez evaluando métodos de aprendizaje en grafos.

Les paso los detalles abajo.



Título: On the Application of Graph Neural Networks for Indoor 
Positioning Systems

Tribunal: Pablo Cancela (IIE-FING), Alberto Castro (IIE-INCO-FING), 
Fernando Gama (Morgan Stanley).

Fecha y hora: miércoles 23 de noviembre a las 10AM

Lugar: laboratorio de software del IIE. Quienes quieran acceder 
virtualmente, se ponen en contacto conmigo.

Resumen: Due to the inability of GPS (Global Positioning System) or 
other GNSS (Global Navigation Satellite System) methods to provide 
satisfactory precision for the indoor localization scenario, indoor 
positioning systems resort to other signals already available on-site, 
typically Wi-Fi given its ubiquity. However, instead of relying on an 
error-prone propagation model as in ranging methods, the popular 
fingerprinting positioning technique  considers a more direct 
data-driven approach to the problem. First of all, the area of interest 
is divided into zones, and then a machine learning algorithm is trained 
to map between, for instance, power measurements from Access Points 
(APs) to the localization zone, thus effectively turning the problem 
into a classification one.

However, although the positioning problem is a geometrical one, 
virtually all methods proposed in the literature disregard the 
underlying structure of the data, using generic machine learning 
algorithms. In this work we consider instead a graph-based learning 
method, Graph Neural Networks, a paradigm that has emerged in the last 
few years and constitutes the state-of-the-art for several problems. 
After presenting the pertinent theoretical background, we discuss two 
possibilities to construct the underlying graph for the positioning 
problem. We then perform a thorough evaluation of both possibilities, 
and compare it with some of the most popular machine  learning 
alternatives. The main conclusion is that these graph-based methods 
obtain systematically better results, particularly with regards to 
practical aspects (e.g. gracefully tolerating faulty APs), which makes 
them a serious  candidate to consider when deploying positioning systems