Active learning in systems of colaborativa filtering / Aprendizagem ativa em sistemas de filtragem colaborativa
AUTOR(ES)
Igor Azevedo Sampaio
DATA DE PUBLICAÇÃO
2006
RESUMO
Nowadays, the amount of available information is much greater than our ability to manage it. There are hundreds of TV channels, dozens of movies on theatres and thousands of products in on-line stores that we can choose from. When we need to take a decision without being able to know all the possible alternatives, a common approach is to rely on recommendations of other people. In the 1990s a new category of computer systems has appeared to automatize the recommendation process. Usually the Recommender Systems, as they came to get known, acquire some indications of user preferences for providing them with a personalized view of the information. One technique that has been widely used in Recommender Systems is the Collaborative Filtering (CF), which uses the similarities among users to make the recommendations. So, for discovering the relevance of a given item i to a target user u, the system relies upon the opinions of other users that have preferences similar to those of the target user u. However a problem that occurs frequently on Recommender Systems is about the arrival of a new user. In that situation, since the newcomer has not rated any information item, the system has no information about his interests. Furthermore the system is not able to generate any recommendation for that user. That problem also occurs on CF-based systems, once the calculation of the similarities among users is based on the items the users have rated in common. A possible solution for diminishing the new user problem is to include an initial preference-acquisition step, where the system presents some items for the new user to rate. However, that must be made in a very efficient way for the system to acquire the most information with a minimum user effort. The machine learning technique where the algorithm controls the presentation order of training examples for optimizing the learning process is called active learning. The application of such technique for improving the acquisition process of user preferences in CF-based systems has been the aim of several studies. In one of the previous work it has been proposed a method called ActiveCP that manages to combine the controversy and popularity of a given information item for deciding the order the items should be presented for receiving user rating. The method has obtained good experimental results. In this work, the use of a new controversy measure is investigated. That measure is capable of solving several restrictions originally present in the proposed methodology for ActiveCP. Furthermore, a new methodology is presented; witch is simpler and has a better applicability for real situations. It keeps the good recommendation results that the original method has obtained. Finally, the new methodology was evaluated in a database of user ratings for movies that emulates the database of a recommender system at startup phase.
ASSUNTO(S)
ciencia da computacao problema do usuÃrio novo, aprendizagem ativa collaborative filtering active learning recommender systems filtragem colaborativa sistemas de recomendaÃÃo new user problem
