Loading...
Please wait, while we are loading the content...
Similar Documents
Matrix factorization models for cross-domain recommendation: Addressing the cold start in collaborative filtering
| Content Provider | Semantic Scholar |
|---|---|
| Author | Tobías, Ignacio |
| Copyright Year | 2017 |
| Abstract | Recommender systems are software tools designed to help users in information access and retrieval tasks. By analyzing previous user interactions with certain information items, these systems estimate the users’ preferences (i.e., tastes, interests and needs) for other items to predict and suggest the most relevant ones. Actively investigated since the nineties, recommender systems have gained popularity and sophistication, and nowadays are essential components of numerous business, education, culture and entertainment services, such as ecommerce sites like Amazon.com and eBay, media content providers like Netflix, YouTube and Spotify, and online social networks like Facebook and Twitter. Multiple recommendation approaches, and remarkably those based on collaborative filtering, have been proposed and successfully implemented over the last years. However, they still have limitations and challenges that in turn represent research opportunities. One of the most notorious of these opportunities is the so called cold start problem, which refers to the situation where a new user has recently registered in a system, and for whom there are not enough preferences to deliver relevant personalized recommendations. Two types of approaches have been explored to address the cold start. The first is represented by techniques that intelligently elicit the preferences from the user, while the second includes methods that make use of additional information to infer user preferences. Within this last type of approaches, cross-domain recommendation has recently emerged as a potential solution, exploiting user preferences and item attributes in domains distinct, but related to the target recommendation domain. Cross-domain recommender systems are currently under research in several fields with particular goals and tasks. In User Modeling, these systems have been proposed as a mechanism to aggregate and mediate user profiles as a cross-system personalization strategy, in Machine Learning they have been explored as a practical application of transfer learning techniques, and in Recommender Systems as a way to mitigate the scarcity of user preference data. This thesis focuses on the study of cross-domain recommender systems as a solution to the cold start. We first provide an in-depth review of the state of the art in the above mentioned research fields, providing a unifying formalization of the problem, and a categorization of existing approaches and evaluation methodologies. We then present three novel adaptations of the matrix factorization technique for cross-domain collaborative filtering. In particular, we propose a number of models that deal with different sources of infor- |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://repositorio.uam.es/bitstream/handle/10486/677710/fernandez_tobias_ignacio_javier.pdf?sequence=1 |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
