融合多因素的兴趣点协同推荐方法研究

Study on Point-of-interest Collaborative Recommendation Method Fusing Multi-factors

兴趣点(Point-of-Interest,POI)推荐是为用户推荐可能感兴趣的地理位置的一项任务,是基于位置社交网络(Location-Based Social Networks,LBSN)服务中的重要研究内容。针对目前POI推荐准确率较低、推荐结果缺乏个性化、情感倾向因素融入差等问题,在综合分析兴趣点的地理位置、分类偏好、流行度、社交与情感倾向等相关影响因素的基础上,提出了融合多因素的兴趣点协同推荐模型(GCSR)。首先,根据POI地理位置数据计算地理相关分数;其次,根据用户的类别偏好,结合POI流行度定义分类偏好分数;然后,根据社交关系计算用户之间的社交关系强度,通过挖掘评论文本计算用户的情感倾向分数,并将二者与协同过滤推荐技术有效结合,从而得到社交情感分数;最后,将地理相关分数、分类偏好分数与社交情感分数有效融合,向用户推荐Top-N兴趣点。在Foursquare真实签到数据集上进行的多组对比实验显示,与基线模型中最好的JRA相比,GCSR模型能够获得更好的推荐效果,准确率和召回率平均提高了1.7%和0.6%。

Point-of-interest (POI) recommendation is a task to recommend geographical locations that users may be interested in.It is an important researches in location-based social networks (LBSN) services.For the existing problems that POI recommendation currently has lower recommendation precision,lacks of personalization in recommendation results,and has poor integration of sentimental orientation factors,etc.,this paper proposed a POI collaborative recommendation model(GCSR) fusing multi-factors based on the comprehensive analysis of POI related influencing factors,such as geographical location,category preference,popularity,social and sentimental orientation and so on.Firstly,the geographical relevance score is calculated based on POI geographical location data.Secondly,category preference score is defined according to users’ category preference and POI popularity.Then,the strength of the social relationships between users is calculated based on the social relationships,the sentimental orientation score of users is calculated by mining the comment text,and the two are effectively combined with the collaborative filtering recommendation technology to obtain the social sentiment score.Finally,geographical relevance score,category preference score and social sentiment score are effectively integrated to recommend Top- N POI.Multiple comparative experiments conducted on Foursquare’s real check-in datasets demonstrate that the GCSR model achieves better recommendation effect,with an ave- rage improvement of 1.7% and 0.6% in precision and recall,compared with the best effective JRA in the baseline models.