基于社交网络的同城活动推荐方法研究

发布时间：2018-05-01 19:27

本文选题：活动社交网络 + 推荐方法　；参考：《西南大学》2017年硕士论文

【摘要】：伴随着互联网的快速发展与互联网技术的不断创新,社交网络日益成熟和完善。在众多的社交网络类型中,有一种以活动为媒介将线上与线下相结合的社交网络——活动社交网络(Event-based Social Network,EBSN)。和传统的社交网络相比,活动社交网络中的用户既可以线上浏览活动信息,又有可以根据活动信息决定是否线下参加该活动。随着时间的推移和网络的发展,活动社交网络中产生的海量数据使得用户难以快速找到自己感兴趣的活动。因此,急需基于活动社交网络的推荐系统来为用户做活动推荐,提高用户体验。社交活动推荐与传统的推荐有所不同,主要有:(1)活动的“一次性消费”特性。活动是人为发起的,具有特定主题、时间、地点,用户只能参加一次,无法像商品一样反复购买,且没有历史评价记录。(2)活动社交网络中有更多的信息可用于推荐。活动社交网络可以形成两种社交关系,一种是用户通过加入兴趣小组等形成的线上社交关系,另一种是用户通过参与相同的社交活动而形成的线下社交关系。此外,还有用户和活动的时间、地理位置等信息。这些不同使得活动推荐不能直接采用传统的推荐方法,因此本文研究社交活动推荐。本文针对上述特点和现有的社交活动推荐中存在不足之处,在已有的推荐相关理论与技术的基础上,给出了本文的基于社交网络的同城活动推荐方法并对其进行有效性验证。本文的主要工作包括:(1)给出了一种基于社交网络的同城活动推荐模型。模型包括数据获取模块、特征提取模块、学习排序模块和推荐模块。数据获取模块解决数据获取问题,并将数据分为训练数据和待推荐数据。特征提取模块是分析数据信息,提取出用户偏好、好友影响、时间匹配度、位置匹配度、活动主题流行度五个特征。学习排序模块是将推荐问题转化为学习排序问题,通过对活动进行学习排序,得到衡量所有特征的最优权重W。推荐模块是根据用户IP判断用户城市,从而选择用户的候选活动,根据最优权重W计算出用户对候选活动的评分,根据评分为用户推荐top-N的活动。(2)分析并提取了用户偏好、好友影响、时间匹配度、位置匹配度、活动主题流行度五个特征,并给出各个特征的计算方法。用户偏好采用基于内容的推荐方法,计算出用户与活动在主题向量的相似度。使用LDA方法表示对用户和活动主题向量,降低了文本维度,缓解了数据稀疏问题。好友影响采用协同过滤方法,将用户偏好视为用户评分,同时将与用户主题相似度最高的K个用户视为其好友。时间匹配度和位置匹配度分别挖掘用户在时间和位置特征上的行为规律,计算用户和活动在时间与位置上的相似度。活动主题流行度这一特征是为了衡量活动主题与城市流行主题之间的相似度,城市流行主题是指该城市近期的参与度最高的M个活动的主题。同时,活动主题流行度可以在一定程度上可以降低冷启动问题对活动推荐的影响。(3)给出了一种基于社交网络的同城活动推荐算法。将活动推荐问题转化为学习排序问题,并借助成对学习排序的思想,将活动组成序列对,分为正序列对和负序列对,从而将问题转化为针对活动序列对的二分分类问题。为综合考虑各个特征的影响,本文对逻辑回归方法进行改进,使其适用于成对学习排序问题。采用平方损失作为损失函数,在求解过程中,采用批梯度下降法进行求解,并为损失函数添加正则化项以防止过拟合,同时添加用户系数以调节用户数据不均衡带来的影响。本文的活动推荐方法是:采用改进的逻辑回归排序方法融合用户偏好、好友影响、时间匹配度、位置匹配度、活动主题流行度五个特征,计算出用户对候选活动的综合评分,并以此进行活动推荐。为验证本文给出的方法的有效性,实验选取准确率和召回率作为推荐结果评估指标,利用豆瓣同城中的数据,与现有的常用的几种活动推荐方法进行对比分析。实验结果表明:相对于单一特征的推荐方法,本文的融合了多特征的活动推荐方法效果更好;相对于其他四种经典的活动推荐方法,本文的改进的逻辑回归排序的活动推荐方法效果更好,能够更有效地为用户进行活动推荐,提高用户的体验,满足用户需求。
[Abstract]:With the rapid development of the Internet and the continuous innovation of Internet technology, social networks are increasingly mature and perfect. In many social network types, there is a social network (Event-based Social Network, EBSN) that combines online and offline with activity as a medium. Users in social networks can not only browse activities online, but also decide whether to take part in the activity according to the activity information. As time goes on and the network develops, the mass data produced in the social network makes it difficult for users to find their own activities quickly. Therefore, the active social network is urgently needed. Recommending systems to recommend activities for users to improve user experience. Social activities recommendation is different from traditional recommendations, mainly: (1) the "one-time consumption" feature of activities. Activities are initiated by people, with specific topics, time, locations, users can only take part in one time, can not be purchased as repeatedly as a commodity, and there is no historical evaluation (2) active social networks have more information to recommend. Active social networks can form two social relationships, one is a user's online social relationship by joining an interest group, the other is a user's social relationship by participating in the same social activity. In addition, there are also users and activities. This paper studies social activities recommendation. This paper, based on the existing recommendation related theories and techniques, presents the social network based on the existing recommendation related theories and techniques. The main work of this paper is as follows: (1) a city activity recommendation model based on social network is given. The model includes data acquisition module, feature extraction module, learning sorting module and recommendation module. Data acquisition module solves the problem of data acquisition and divides data into training. The feature extraction module is the analysis of data information, which extracts five characteristics: user preference, friend influence, time matching degree, position matching degree, and activity theme popularity. The learning sorting module is to transform the recommendation problem into a learning sort problem. The optimal weight W. recommendation module is based on the user's IP to judge the user's city, and then selects the user's candidate activities, calculates the user's score on the candidate activity according to the optimal weight W, and according to the score the user recommends the activity of the top-N. (2) analyze and extract the user preference, the friend influence, the time matching degree, the position matching degree, the activity theme popularity five The user preference uses a content based recommendation method to calculate the similarity between the user and the activity in the subject vector. Using the LDA method to represent the user and the activity theme vector, the text dimension is reduced and the data sparsity is alleviated. The friend influence uses the collaborative filtering method to use the user preferences. The K users with the highest similarity of the user's theme are considered as their friends. The time matching and position matching degree is used to discover the behavior rules of the user on the time and position characteristics, and calculate the similarity between the user and the activity in the time and position. The feature of the activity topic flow degree is to measure the activity theme and the activity topic. Similarity between urban popular themes, urban popular theme refers to the theme of the city's most recent M activities. At the same time, activity theme popularity can reduce the impact of cold start problems on activities recommendation. (3) a city based activity recommendation algorithm based on social network is given. The problem of recommendation is transformed into a learning sort problem, and by means of the idea of a pair learning sort, the activity is composed of sequence pairs, which are divided into positive sequence pairs and negative sequence pairs, and then the problem is converted into a two classification problem aiming at the sequence pairs of activity. In order to consider the influence of each characteristic, the paper improves the logic regression method to make it suitable for the pair. We use the square loss as the loss function. In the process of solving the problem, the batch gradient method is used to solve the problem, and the regularization term is added to the loss function to prevent the overfitting. At the same time, the user coefficient is added to adjust the influence of the user's data imbalance. In order to verify the validity of the method given in this paper, the accuracy and recall rate are selected as the evaluation index of the recommended results. The order method combines five characteristics of user preference, friend influence, time matching degree, position matching degree and activity theme popularity. The data in the same city are compared with the existing methods of active recommendation. The experimental results show that the combination of the multi feature recommendation method is better than the single feature recommendation method. Compared with the other four classical methods of activity recommendation, the improved logical regression sequencing is used in this paper. The recommendation method has better effect, can more effectively recommend activities for users, improve user experience and meet user needs.

【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】