본문 바로가기
  • Home

The Study of Recommendation Algorithm's Predictive Accuracy Improvement Using Structural Holes on Trust-based Social Networks

  • Journal of Knowledge Information Technology and Systems
  • Abbr : JKITS
  • 2017, 12(1), pp.209-217
  • DOI : 10.34163/jkits.2017.12.1.019
  • Publisher : Korea Knowledge Information Technology Society
  • Research Area : Interdisciplinary Studies > Interdisciplinary Research
  • Published : February 28, 2017

Kang, BooSik 1

1목원대학교

Accredited

ABSTRACT

Improving predictive accuracy of recommendation algorithms is a major work in the area of recommender systems. Collaborative filtering is the most popular method for product recommender systems. User-based collaborative filtering recommends products using the information about product preference of Neighbors. Recently, some studies enhancing predictive accuracy of recommender systems using information about the relationship of friend or trust between users has been published. This study proposed a method constructing trust-based social networks and appling structural holes for improving predictive accuracy of collaborative filtering. It constructs a social network using dataset represented trust relationship and finds user impact using structural holes that is one of methods for social network analysis. Neighbor's similarities of a target user for recommendation are adjusted by neighbor's impact which was found in before procedure. This study experimented two techniques for adjusting similarities. LinearImpCF adjusts neighbor's similarities multiplying by and neighbor's impact. ExpImpCE adjusts neighbor's similarities multiplying neighbor's impact to the power. To validate, the proposed algorithms were applied to filmtrust dataset. The results of 10-fold cross validation showed that mean MAEs of LinearImpCF and ExpImpCF were lower than mean MAE of conventional CF. We knew that the proposed method improved the predictive accuracy slightly. To test statistical significance, we experimented 10-fold cross validation repeatedly three times. We confirmed the statistical significance by paired t-test using experiment results. In conclusion, we knew that the proposed recommendation algorithm combined collaborative filtering and user's impact by structural holes on trust-base social networks between users improved the predictive accuracy.

Citation status

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