Aprendizado neural de representação de conteúdo para sistema de recomendação de filmes

Abstract

Recommender systems are software used to generate personalized lists according to users profiles. The area is new and is growing since the internet popularization having its roots in information retrieval. Collaborative filtering is the most common approach of recommender systems used in both academy and industry because content-based filtering has problems such as lack of semantic information and poor content extraction techniques from items. Nowadays there are more content available in the form of multimedia such as video, images and text. Also, there are advances in pattern recognition though techniques like convolutional neural networks. In this work a convolutional neural network is used to extract features from movie trailers frames to further use these features to create a content-based recommender system with the goal of assessing whether the success of such networks on tasks like image classification and object detection also occur in the recommendation context. To evaluate that, the proposed method was compared with a media aesthetic detection method, two methods of feature extraction from text using TF-IDF and the traditional user and item collaborative filtering methods. Our results indicate that the proposed method is superior to the other content-based methods and is competitive to the collaborative filtering methods, being superior to the item-collaborative method regarding classification accuracy, and being superior to all other methods regarding execution time. In conclusion, we can state that the method using convolutional neural networks to represent items is promising for the recommender systems context.