【作者】 李松；

【导师】 彭蓉；

【作者基本信息】 武汉大学 ， 软件工程， 2021， 硕士

【摘要】 随着移动互联网的迅速发展,一些基于位置的社会网络(Location-based Social Network,LBSN)平台日益普及,这些LBSN平台的用户在地理位置点上签到、分享位置以及活动信息等行为,在平台产生了大量的签到数据。这些数据催生出一项新的推荐服务——基于位置的推荐服务。在地理信息系统中,餐厅、影院、体育馆等具有唯一标识、且反映用户偏好的位置被称为兴趣点(Point-of-Interest,POI),因此,位置推荐又被称为POI推荐。用户在某一个兴趣点可以轻松的分享与位置关联的内容,这些内容很多都可以被用于改善LBSN平台用户体验的兴趣点推荐中。其中,兴趣点的位置、访问兴趣点的时间、频次等是最常用的兴趣点推荐因素。但是,大多数方法在分析时间因素对用户兴趣点推荐产生的影响时,要么捕捉该时间特征时的粒度比较粗糙,方法比较传统,无法捕捉高维特征;要么仅从签到序列的近期签到记录出发,没有考虑到用户长期的签到序列所形成的时间周期性特征在时间点的差异性上对其偏好所产生的影响。针对以上问题,本文设计了一种基于GRU模型的兴趣点推荐模型—Time4Pre。该模型首先在数据预处理阶段对用户签到时间进行编码生成时间戳,在此基础上为兴趣点生成特征时间戳,再利用兴趣点的特征时间戳生成了高质量的负样本序列,然后基于GRU模型捕捉用户的序列偏好和时间偏好。对序列偏好的捕捉是采用了GRU作为基本模型,而对时间偏好的捕捉是以两个签到记录的时间戳间隔为基础,充分的考虑了用户长期的签到序列所形成的时间周期性特征对其偏好所产生的影响,最后,从线性和非线性两个角度对序列偏好和时间偏好进行融合,为用户进行地理兴趣点的推荐。本文具体的贡献如下:(1)从时间影响和序列影响两个角度详细分析了两个公开的签到数据集。得出以下结论:在时间影响上,1)用户在工作日上的签到规律性要强于周末;2)无论是工作日还是周末,任意两次签到记录时间间隔的规律性要弱于连续两次签到记录;3)用户更倾向于去与已经签到过的记录时间间隔较小的兴趣点签到,并且在连续签到记录上这种时间关联要更为紧密。以上发现表明用户的签到行为与签到记录的时间间隔是有明显关系的。在序列影响上,用户的连续性签到行为表现出明显的序列性特征,用户的多次连续性签到POI记录会对用户的签到行为产生影响,其中连续两次连续性签到POI记录对用户签到行为产生的影响最大。(2)提出了一种基于兴趣点特征时间戳的负样本生成方法—时间特征法。Time4Pre模型在数据预处理阶段,在利用用户签到时间点生成时间戳的基础上,为兴趣点生成特征时间戳,再利用兴趣点的特征时间戳构建用户签到记录的负样本序列。实验结果表明,这种构建负样本序列的方法在提高负样本序列质量的同时还能保证训练时的效率,并且能够提高模型的推荐效果。(3)捕捉用户的时间偏好和序列偏好并对两者进行融合。不同于以往对时间周期性特征的模糊化处理,Time4Pre模型以反映签到时间差异性的签到记录时间戳间隔为基础,区别化对待周末和工作日,在捕捉用户的时间偏好的同时捕捉用户的序列偏好,然后使用线性和非线性的两种方式融合这两种偏好。线性方式参数比非线性方式更为简单,但非线性方式能更好的适应复杂的应用场景。本文基于两个公开的签到数据集,设计了大量的实验验证了所提出的模型的有效性,并探究了不同的影响因素对实验效果的影响。实验表明,捕捉用户时间偏好的方法对于提升模型的推荐效果非常有效,两种融合方式的实验效果都优于目前最好模型。更多还原

【Abstract】 With the rapid development of the mobile Internet,some Location-based Social Network(LBSN)platforms are becoming more and more popular.Users of these LBSN platforms will check in at geographic locations,share location,and activity information,which will generate a large amount of check-in data.These data can give birth to a new recommendation servicelocation recommendation.In a geographic information system,locations that have unique identifications such as restaurants,theaters,and stadiums that reflect user preferences are called points-of-interest(POI).Therefore,location recommendation is also called POI recommendation.Users can easily share location-related content at a certain point of interest,and many of these content can be used in point-of-interest recommendations to improve the user experience of the LBSN platform.Among the existing points of interest recommendation methods,there is a lack of effective methods for constructing high-quality negative samples of user check-in records.When discussing the impact of time factors on user points of interest recommendation,either the granularity of capturing the time feature is relatively rough,the method used is relatively traditional,and high-dimensional features cannot be captured;or it is only based on the recent check-in records of the check-in sequence,without considerating the time cyclical characteristics formed by the long-term check-in sequence of the user having an impact on their preference in terms of the difference in time.In response to the above problems,this paper designs a point-of-interest recommendation model based on the GRU model—Time4Pre.The model first encodes the user check-in time to generate a time stamp in the data preprocessing stage,and then generates a characteristic time stamp for the point of interest,and generates a high-quality negative sample sequence using the characteristic time stamp of the point of interest.Then it captures the user’s sequence preference and time preference based on the GRU model.The capture of sequence preference uses GRU as the basic model,and the capture of time preference is based on the time stamp interval of two check-in records,fully taking into account the time periodic characteristics formed by the user’s long-term check-in sequence having an impact on their preference in terms of the difference in time.Finally,the sequence preference and time preference are merged from both linear and non-linear perspectives to recommend geographic points of interest for users.The specific contributions of this article are as follows:(1)Two public check-in record data sets are analyzed in detail from the perspectives of time influence and sequence influence,which can draw the following conclutions:In terms of time influence,1)the regularity of the user’s check-in records on workdays is stronger than that on weekends;2)the regularity of the time interval between any two checkin records is weaker than that of two consecutive check-in records,whether it is a workday or a weekend.3)users are more inclined to check-in with points of interest that have been checked-in with a shorter time interval,and this time correlation should be closer in continuous check-in records.The above findings indicate that there is an obvious relationship between the user’s check-in behavior and the time interval of the check-in record.In terms of sequence influence,the user’s continuous check-in behavior shows obvious sequential characteristics.The user’s multiple consecutive check-in POI records will have an impact on the user’s check-in behavior,and two consecutive check-in POI records have the greatest impact on the user’s check-in behavior.(2)A method of generating negative samples based on the characteristic timestamp of interest points—Time character method is proposed.In the data preprocessing stage of the Time4 Pre model,based on the time stamp generated by the user’s check-in time,the characteristic time stamp is generated for the points of interest,and then the characteristic timestamp of the point of interest is used to construct a negative sample sequence of the user’s check-in record.Experimental results show that this method can not only improve the quality of negative sample sequences,but also ensure the efficiency of negative sample sequences when used for training,which is also helpful to improve the recommendation effect of the model.(3)Capturing the user’s time preference and sequence preference and integrate the two.The user’s check-in record has obvious time periodic characteristics.Different from the previous fuzzification of periodic characteristics of time,the Time4 Pre model is based on the check-in record timestamp interval reflecting the difference of check-in time,treats weekends and working days differently,and captures the user’s sequence preference while capturing the user’s time preference,and then use linear and non-linear methods to fuse these two preferences.Linear method’s parameters are simpler than nonlinear methods,but non-linear method can better adapt to complex application scenarios.Based on two public check-in data sets,this paper designs a large number of experiments to verify the effectiveness of the proposed model,and explores the influence of different influencing factors on the experimental results.Experiments show that the method of capturing user time preference is very effective in improving the recommendation effect of the model,and the experimental effects of the two fusion methods are better than the best model currently.更多还原