학술논문
심층신경망을 이용한 협업필터링 방법 연구 / A Study on Collaborative Filtering using Deep Neural Network
Document Type
Dissertation/ Thesis
Author
Source
Subject
Language
Korean
Abstract
본 논문에서는 협업필터링 방법에 의한 추천시스템 구현을 위하여, 심층신경망을 이용하는 ‘임베딩벡터유사도’(SEV) 방법을 제안한다. SEV 방법은 사용자들의 특성을 임베딩벡터로 변환하고, 변환된 임베딩벡터의 유사성에 근거하여, 참조사용자들을 선정하고, 참조사용자들의 사용 이력이 높은 상품을 추천하는 방법이다.심층신경망을 이용한 협업필터링 방법으로는 He 등(2017)이 제안한 ‘신경망행렬분해’(NeuMF) 방법이 널리 알려져 있다. 본 논문에서는 무비렌즈데이터에 적용하는 경우에서, SEV 방법과 NeuMF 방법의 성능 지표를 계산하고 비교하였다. 본 논문에서 제안한 SEV 방법이 NeuMF 방법에 비하여 최소한 필적할만하거나 더 우수한 성능을 보였다.임베딩벡터유사도 방법은 평가 자료의 (사용자, 상품) 공간에서의 희박성에 영향을 받지 않고 안정적으로 적용할 수 있는 알고리즘으로, 향후 다양한 상품의 추천시스템에 적용될 수 있을 것이라 예상된다.
In this paper, a new algorithm, to implement a collaborative filtering method as a recommendation system, is proposed. The new algorithm adopts deep neural net method to capture the user characteristics into embedding vectors. The new algorithm, called SEV method in short, uses the similarity of embedding vectors (SEV) of users to find the reference users for each target user. The top-k frequently referenced items by the reference users are recommended to the target user. He, et al. (2017) proposed a collaborative filtering method using deep neural network, called neural network matrix factorization method, NeuMF method in short. In this paper, the performance indices of SEV method and NeuMF method were evaluated in the case when the methods were applied to MovieLens 100K data. The SEV method proposed in this paper showed the performance at least comparable with or better than the NeuMF method.The SEV method is an algorithm that can be applied stably without being affected by the sparsity of the rating data in the (user, item) space. The SEV method is expected to be applied widely to the recommendation systems of various products in the future.
In this paper, a new algorithm, to implement a collaborative filtering method as a recommendation system, is proposed. The new algorithm adopts deep neural net method to capture the user characteristics into embedding vectors. The new algorithm, called SEV method in short, uses the similarity of embedding vectors (SEV) of users to find the reference users for each target user. The top-k frequently referenced items by the reference users are recommended to the target user. He, et al. (2017) proposed a collaborative filtering method using deep neural network, called neural network matrix factorization method, NeuMF method in short. In this paper, the performance indices of SEV method and NeuMF method were evaluated in the case when the methods were applied to MovieLens 100K data. The SEV method proposed in this paper showed the performance at least comparable with or better than the NeuMF method.The SEV method is an algorithm that can be applied stably without being affected by the sparsity of the rating data in the (user, item) space. The SEV method is expected to be applied widely to the recommendation systems of various products in the future.