Collaborative Ranking with triple pairwise constraints for Top-K Location Recommendation

Abstract

본 연구는 세가지 쌍방조건식을 활용하여, Top-K 추천을 위한 협업 필터링을 활용한 랭킹 모델을 설계하였다. 첫째, 기존 연구에서 implicit data에서 관측되지 못한 값 또는 결측 값을 활용하지 못하였다. 제안 된 모델은 사용자가 가지지 못한 정보들을 사용자의 친구의 방문 정보를 활용하여, 학습 단계에 적용되었다. 두번째로 각 사용자의 친구들이 특정 장소에 방문을 한 총 인원수가 많을수록, 해당 장소들이 상위 K개의 추천리스트에 반영이 될 수 있도록 학습을 하고자 하였다. 그 결과 기존의 모델보다 추천 성능이 향상됨을 실험을 통해 입증하였다.

In this thesis, we propose a collaborative ranking model for Top-K recommendation (SFBPR) by injecting social network preference. We also included a weight variable that represents the frequency of the places visited by the user’s friends and utilize triple pairwise constraints in learning step, so that the locations that have the highest number of visits would be located higher in the recommendation list. We utilize and set triple pairwise constraints and assumption. The first assumption is that the location where the user(u) visited is prior the the place visited by the friend, and second assumption is that the location that the friend of user(u) visited is higher in rank than the location that user(u) did not visited. Thanks to exploit social preference, latent facotrs in latent space, SFBPR achieved higher ranking metrics such as precision, recall, NDCG , AUC score than did existing models. Our experimental results show that SFBPR is proper method for personalized POI recommendation task based on two real-world dataset. SFBPR model’s performance. The SFBPR model improves the recommendation performance by 10% in NDCG, Recall score and 6% in precision score when compared with the performance of the existing competing models.