Exploring attributes, sequences, and time in recommender systems: from classical to point-of-interest recommendation

• Autores: Pablo Sánchez
• Directores de la Tesis: Alejandro Bellogín Kouki (dir. tes.), Pablo Castells Azpilicueta (tut. tes.)
• Lectura: En la Universidad Autónoma de Madrid ( España ) en 2021
• Idioma: español
• Tribunal Calificador de la Tesis: Tommaso Di Noia (presid.), Juan de Lara Jaramillo (secret.), Ludovico Boratto (voc.)
• Programa de doctorado: Programa de Doctorado en Ingeniería Informática y de Telecomunicación por la Universidad Autónoma de Madrid
• Materias:
  • Matemáticas
    • Ciencia de los ordenadores
      • Inteligencia artificial
      • Informática
• Enlaces
  • Tesis en acceso abierto en: Biblos-e Archivo
• Resumen

  • Since the emergence of the Internet and the spread of digital communications throughout the world, the amount of data stored on the Web has been growing exponentially. In this new digital era, a large number of companies have emerged with the purpose of filtering the information available on the web and provide users with interesting items. The algorithms and models used to recommend these items are called Recommender Systems. These systems are applied to a large number of domains, from music, books, or movies to dating or Point-of-Interest (POI), which is an increasingly popular domain where users receive recommendations of different places when they arrive to a city.

    In this thesis, we focus on exploiting the use of contextual information, especially temporal and sequential data, and apply it in novel ways in both traditional and Point-of-Interest recommendation. We believe that this type of information can be used not only for creating new recommendation models but also for developing new metrics for analyzing the quality of these recommendations. In one of our first contributions we propose different metrics, some of them derived from previously existing frameworks, using this contextual information. Besides, we also propose an intuitive algorithm that is able to provide recommendations to a target user by exploiting the last common interactions with other similar users of the system.

    At the same time, we conduct a comprehensive review of the algorithms that have been proposed in the area of POI recommendation between 2011 and 2019, identifying the common characteristics and methodologies used. Once this classification of the algorithms proposed to date is completed, we design a mechanism to recommend complete routes (not only independent POIs) to users, making use of reranking techniques. In addition, due to the great difficulty of making recommendations in the POI domain, we propose the use of data aggregation techniques to use information from different cities to generate POI recommendations in a given target city.

    In the experimental work we present our approaches on different datasets belonging to both classical and POI recommendation. The results obtained in these experiments confirm the usefulness of our recommendation proposals, in terms of ranking accuracy and other dimensions like novelty, diversity,and coverage, and the appropriateness of our metrics for analyzing temporal information and biases in the recommendations produced.