一种基于矩阵分解的上下文感知POI推荐算法

Context-Aware POI Recommendation Based on Matrix Factorization

近年来,随着移动设备的普及,基于位置的社交网络(Location-Based Social Network,LBSN)逐渐被人们广泛使用并成为一种新型的社交媒体.LBSN能够记录丰富的上下文信息,例如用户社交网络、POI地理位置、POI类别信息等,这无疑为个性化的POI(Point-of-Interest)推荐系统带来了巨大的发展机遇.但是如何建模这些上下文信息对POI推荐的影响并将它们有效地融合成为了一大难点,另外用户签到数据的稀疏性也为POI推荐带来巨大的挑战.为了克服上述挑战,该文提出了一个基于矩阵分解的上下文感知POI推荐模型.具体地,该文从多个方面建模用户的签到行为,除了利用用户的签到数据,还考虑了POI的地理位置对用户签到行为的影响,用户更愿意访问那些距离近并且符合自身偏好的POI.另外,为了进一步缓解签到数据的稀疏性,该文还利用了用户社交网络数据和POI类别信息.最后,该文提出了一个通用的矩阵分解模型,它能有效地融合上述上下文信息,并且具有良好的可扩展性和较低的时间复杂度.在两个真实的LBSN数据集上的实验结果表明,该文提出的方法在推荐的准确性上远优于当前流行的POI推荐算法.

With the popularity of mobile devices,Location-based Social Network(LBSN)has been widely used and becomes a new form of social media in recent years.LBSN can record rich context information,such as social networks,geographical information,POI category information,etc.The context information provides a great opportunity to build personalized POI(Point-of-Interest)recommender systems.In POI recommendation,users’ check-in behavior is not only affected by the users’ own preferences,but also influenced by various context information in the surrounding environments.Therefore,how to model the influence of the context information on the users’ check-in behavior and effectively integrate them with the users’ preferences becomes a major difficulty.In addition,the number of users and POIs in the LBSN often reaches millions,which makes the user-POI check-in matrix very large and sparse.The sparsity of user’s check-in data also poses a huge challenge for POI recommendation.In this paper,we propose a context-aware POI recommendation model based on matrix factorization,which addresses the above challenges at some extent.Specifically,we attempt to model the users’ check-in behavior from multiple aspects.Firstly,we use matrix factorization technology to learn users’ own preferences from the users’ check-in data(U),and consider the influence of POI category information(C)on the users’ preferences,since users often prefer a certain type of POI rather than a specific POI.Secondly,POI’s geographical location(G)has a great impact on user’s check-in behavior.Users prefer to visit POIs that are closer and meet their own preferences.We use kernel functions to model the distance distribution between any POI pair,and then use item based collaborative filtering to calculate the users’ preferences for POIs in terms of geographical location.Thirdly,taking into account that user’s check-in behavior may be influenced by friends,we use user-based collaborative filtering to model the user social network(S)on the users’ check-in behavior,and further relieve the sparsity of the check-in data.Finally,we propose a general matrix factorization model UCGSMF.When modeling users’ check-in behavior with the influence of context information and their own preferences,different context information strategies are applied to visited POIs and non-visited POIs.In this way,users’ own preferences can be better fitted.What’s more,the model has good extensibility,and it is very flexible in the modeling of context information.At the same time,by adopting an improved alternating least-squares algorithm,the model has a lower time complexity.In this paper,a large number of experiments are conducted on Dianping and Foursquare datasets.First,we analyze the recommendation performance under different algorithms.Precision and recall are used to evaluate the performance of the algorithms.Experimental results show that the recommendation performance of our model is much better than the state of the art POI recommendation algorithms.Then,considering the important influence of context information on POI recommendation,we analyze the performance of different context information.The experimental results show that the POI category information factor can indeed improve the POI recommendation performance when a suitable value is obtained.Compared with social network information,geographical location information has higher impact on POI recommendation.Finally,we compare the training time of different algorithms.The experimental results verify that our model has a lower time complexity.