Desarrollo de sistemas de recomendación usando información de redes sociales para paliar el problema de arranque en frío

Abstract

Nowadays we live in a period where there are plenty of options for consuming content online, either books, films or music. New material is released every day and users can consume this content with just a couple of clicks. Despite of this vast amount of options, or maybe due to that, it is more difficult than ever for users to find content that they would enjoy consuming. Sometimes this process can feel like looking for a needle in a haystack. The role of recommender systems is to filter all this content and to provide only the interesting items to the users. These systems are normally based on historical data from the users with other items. For example previous ratings of items can be used to recommend similar items to the ones were ranked highest. One of the most common pitfalls from these systems is the cold start problem. This problem occurs when either a new user or new item is introduced in the system and, therefore, there are no previous data that could be leverage by the recommender systems in order to create recommendations. This problem has an ever-growing importance due to the huge offer of online services for consuming content. These systems need to be prepared to engage users that recently join their platforms by offering them contents that the users would enjoy. Otherwise there is a high risk that these users would leave and find another platform. This topic is widely studied in the literature but due to its importance and its peculiarities, like different domain behaviour or lack of appropriate datasets to study this problem, there is still much to study and research and the current state-of-the-art algorithms have room for improvement. In this proposal, this issue is addressed from different perspectives and applied for different domains and scenarios. The goal of this work is to alleviate the cold start problem and for that we develop models using artificial intelligence algorithms that make use of users’ contextual data from their social media profiles. These models outperform the state-of-the-art models for cold start problem. In addition to that, in this proposal a dedicated dataset has been created. This dataset is optimised for the study of the cold start problem and has data about movie rating, movie description and user description. This will ease future researches since such datasets are rather scarce in the literature. Other areas of this proposal is hedge fund management or reinforcement for recommendation of education resources in the field of oral surgery and implantology and how to address these topics when there is not much previous data available. This work obtained optimal results, matching, and even improving results from state-of- the-art algorithms for recommendation systems with cold-start problem. This is obtained through the leverage of implicit data, extracted automatically from different sources not having to require the user to provide any manual data.