본문 바로가기
  • Home

Improving Accuracy of Movie Recommender System Using Word2Vec and Deep Neural Networks

  • Journal of Knowledge Information Technology and Systems
  • Abbr : JKITS
  • 2018, 13(5), pp.561-568
  • DOI : 10.34163/jkits.2018.13.5.006
  • Publisher : Korea Knowledge Information Technology Society
  • Research Area : Interdisciplinary Studies > Interdisciplinary Research
  • Published : October 31, 2018

Kang, BooSik 1

1목원대학교

Accredited

ABSTRACT

Word2Vec is a most popular method in text mining area, recently. It converts words to vectors using association among words in sentences. Similar words are nearly located in the vector space. As deep learning technology has developed rapidly, deep neural network that adopts deep learning is being applied in many areas. Improving predictive accuracy of recommender algorithms is a major work in the area of recommender systems. This study proposed an integrated method for movie recommender systems using Word2Vec and deep neural networks. First, it constructs Corpus of users and movies. It generates sentences for constructing Corpus. It finds users that give same rating to a movie, generates a sentence with those users, and constructs the Corpus of users using the sentences. It finds movies that are given same rating from an user, generates a sentence with those movies, and constructs the Corpus of movies using the sentences along the same lines. Secondly, it calculates user vectors and movie vectors using Word2Vec. Thirdly, it learns a model of deep neural networks with inputs composed of the user vectors and the movie vectors. Lastly, it recommends movies to the target user through the learned deep neural networks. To validate, the proposed method was applied to filmtrust dataset. The experimental results of 10-fold cross validation showed that the proposed method improved accuracy greatly than conventional collaborative filtering method(uCF). Also, it showed that the proposed method could solve problems with comparatively high accuracy, recommendation problems to new customers and new products, but the conventional collaborative filtering methods had difficulties recommendation to those problems.

Citation status

* References for papers published after 2023 are currently being built.