统计与数据科学学院知识库

随着移动互联网的飞速发展，移动终端设备已经完全融入到我们的日常生活、工作学习、娱乐以及社交活动中，随之而来的是用户难以从海量的行为信息中获取有效资源。如何挖掘移动终端用户的网络行为并且进行精准的个性化推荐服务已成为运营商亟待需要解决的问题，然而传统的用户行为分析从单一场景出发，未考虑用户的行为习惯可能由于时间因素或者外部环境的变化发生改变，导致个性化推荐的信息不符合用户实际的行为习惯。

因此，本文从上网时段、访问内容两方面来分析移动终端用户的网络行为特征，并且利用BP神经网络优化协同过滤推荐算法中基于平均分的预测评分公式，对于运营商而言，可以利用预测结果为用户制定个性化网络服务的推荐列表，有助于提升用户对网络应用的黏度和满意度。为了实现移动终端用户网络行为特征的挖掘，本文从用户的上网时段、访问内容两方面进行分析。首先，选用欧氏距离度量用户上网时间的相似性，通过验证发现用户上网时间在最大上网时段处具有相似性规律，基于此利用层次聚类挖掘用户上网时间的4种行为模式；其次，采用k-means聚类算法挖掘用户访问内容的7种行为模式；最后关联用户上网时段的4种行为模式和访问内容的7种行为模式，得到用户在不同时间段的上网行为特征。

为提高移动终端用户的上网体验，使其实现更加精准的个性化网络服务。本文利用TF-IDF算法计算用户对网络服务类别的喜好程度，针对传统推荐算法的预测评分准确性较低的问题，本文将BP神经网络与协同过滤推荐算法相结合来优化预测评分，并将其与改进的协同过滤推荐算法、基于情感分析的协同过滤推荐算法进行对比。研究结果表明：通过BP神经网络改进协同过滤推荐算法，其均方根误差和平均绝对误差均小于改进相似度的协同过滤推荐算法、基于情感分析的协同过滤推荐算法，因此本文所使用的个性化推荐算法能够更有效地提高评分预测的准确性且得到了更好的效果。最后提出个性化推荐算法在广告精准投放、网络信息精准推送与电子商务营销三个方面的应用价值。

With the rapid development of the mobile Internet, mobile terminal devices have been fully integrated into our daily life, work, study, entertainment and social activities, and it is difficult for users to obtain effective resources from massive behavioral information. How to mine the network behavior of mobile terminal users and provide accurate personalized recommendation services has become an urgent problem for operators to solve. However, the traditional users behavior analysis starts from a single scenario and does not consider the user's behavioral habits due to time factors or external environment. changes, resulting in personalized recommendations that do not conform to the actual behavior of users.

Therefore, this thesis analyzes the network behavior characteristics of mobile terminal users from the aspects of Internet access time and access content, and uses the back propagation neural network to optimize the average score-based prediction scoring formula in the collaborative filtering recommendation algorithm. For operators, the prediction results can be used. Developing a recommendation list of personalized network services for users can help improve users' stickiness and satisfaction with network applications. In order to realize the mining of mobile terminal users' network behavior characteristics, this thesis analyzes the users' online time period and access content. First, the Euclidean distance is used to measure the similarity of users' online time. Through verification, it is found that the users' online time has a similarity law at the maximum online period. Based on this, hierarchical clustering is used to mine four behavioral patterns of users' online time. Second, k-means clustering algorithm mines 7 behavior patterns of users accessing content; finally, correlates the 4 behavior patterns during the users' surfing period with the 7 behavior patterns of accessing content, and obtains the surfing behavior characteristics of users in different time periods.

In order to improve the Internet experience of mobile terminal users and enable them to achieve more accurate personalized network services. In this thesis, the termfrequency–inversedocumentfrequency algorithm is used to calculate the users’ preference for the network service category. In view of the low accuracy of the prediction score of the traditional recommendation algorithm, this thesis combines the back propagation  neural network with the collaborative filtering recommendation algorithm to optimize the prediction score, and use it to optimize the prediction score.Compared with  improved collaborative filtering recommendation algorithm and the collaborative filtering recommendation algorithm based on sentiment analysis. The research results show that the root mean square error and mean absolute error of the collaborative filtering recommendation algorithm improved by the back propagation neural network are smaller than those of the collaborative filtering recommendation algorithm based on the improved similarity and the collaborative filtering recommendation algorithm based on sentiment analysis. Therefore, the personalized recommendation algorithm used in this thesis is the recommendation algorithm can improve the accuracy of rating prediction more effectively and get better results. Finally, the application value of personalized recommendation algorithm in three aspects: advertising precision placement, network information precision push and e-commerce marketing is proposed.