基于Spark平台的通信网用户流失预测研究

发布时间：2018-01-07 09:17

本文关键词：基于Spark平台的通信网用户流失预测研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：近几年,随着移动通信技术的高速发展,移动通信网的通信用户数量急剧增长,通信市场接近饱和,电信运营商之间的竞争也日益激烈。同时,已经接近饱和的市场使得运营商们更加关注用户资源流失的问题。对于通信运营商而言,通过利用通信用户使用移动终端时产生的各种数据,可以预测出潜在的流失用户,并对这些潜在的流失用户进行挽留,从而可以维持市场份额和利润。因此,用户流失预测问题的研究对于通信运营商而言,有着重要意义。本文从神经网络算法的训练速度和特征的选取两个方面来对通信网用户流失预测问题进行深入研究。首先针对BP(Back Propagation)神经网络算法进行研究。BP神经网络算法有两种权重更新策略,单样本(Individual)更新和批样本(Full-Batch)更新。对于Full-Batch BP神经网络算法,每更新一次权重需要对样本集中所有的样本都进行计算,因此,该算法的训练过程非常耗时,但是它可以并行实现。对于Individual BP神经网络算法,每更新一次权重仅需要对样本集中的某一个样本进行计算,因此,该算法的权重更新很快,但它不能并行实现。通过结合Full-Batch BP和Individual BP两种神经网络算法,我们提出了基于Spark平台的Mini-Batch BP神经网络分布式算法来改善传统的两种BP神经网络算法的性能。实验证明,和Full-Batch BP神经网络算法相比,Mini-Batch BP神经网络分布式算法在不失预测准确率的情况下,其训练时间大大减小。接下来,对Mini-Batch BP神经网络分布式算法的参数K的取值进行探讨。通过实验发现K的取值对训练时间有较大影响,并且得到了一个针对最优的K的取值的定性的结论。然后针对通信网用户流失预测问题的特征选取进行研究。首先提取了七个特征,其中六个用户通话行为特征和一个用户相关性特征。用户相关性特征是流失用户给他的邻居用户带来的影响。我们使用激活扩散算法去提取这个特征。随后,分别使用前六个通话行为特征和所有的七个特征来作为训练集进行训练并预测。实验结果表明,在加上了一个用户相关性特征后,预测性能会有所提升。接下来,从对用户之间的相关性统计和特征的相对重要性两个方面证实了用户相关性特征的作用。最后,通过结合两个研究点,针对用户流失预测问题提出了一个快速精确的预测模型。本论文中使用的Mini-Batch BP神经网络分布式算法可以在实际训练和预测的过程中加快速度,帮助快速及时的预测出流失用户。同时,用户相关性的特征也能有效提高预测精度。因此,本文提出的快速精确的预测模型对于流失用户预测的实际应用场景具有重要意义。
[Abstract]:In recent years, with the rapid development of mobile communication technology, the number of communication users in mobile communication networks has increased dramatically, the communication market is nearly saturated, and the competition between telecom operators is becoming increasingly fierce. The already saturated market makes operators pay more attention to the problem of the loss of user resources. For communication operators, the data generated by using mobile terminals through the use of communication users. The potential loss of users can be predicted, and these potential lost users can be retained to maintain market share and profits. This paper studies the problem of user churn prediction in communication network from two aspects of training speed and feature selection of neural network algorithm. Back propagation neural network algorithm for research. BP neural network algorithm has two weight updating strategies. Single sample individual update and batch sample Full-Batch update. For Full-Batch BP neural network algorithm. Every update of the weight needs to calculate all the samples in the sample set, so the training process of the algorithm is very time-consuming. However, it can be implemented in parallel. For the Individual BP neural network algorithm, only one sample in the sample set needs to be calculated for each update of the weight. The weight of the algorithm is updated quickly, but it can not be implemented in parallel, by combining Full-Batch BP and Individual BP neural network algorithm. We propose a Mini-Batch BP neural network distributed algorithm based on Spark platform to improve the performance of two traditional BP neural network algorithms. Compared with Full-Batch BP neural network algorithm, the distributed algorithm of Mini-Batch BP neural network has no loss of prediction accuracy. The training time is greatly reduced. Then, the parameter K of the distributed algorithm of Mini-Batch BP neural network is discussed. It is found that the value of K has a great influence on the training time. A qualitative conclusion is obtained for the optimal K value. Then the feature selection of the communication network user churn prediction problem is studied. Firstly seven features are extracted. Among them, six call behavior features and one user correlation feature. The user correlation feature is the influence of the lost user on his neighbor user. We use the activation diffusion algorithm to extract this feature. The first six call behavior features and all seven features are used as training sets for training and prediction. The experimental results show that after adding a user correlation feature. The prediction performance will be improved. Then, from the two aspects of the correlation statistics and the relative importance of the feature, the function of the user correlation feature is confirmed. Finally, by combining the two research points. This paper presents a fast and accurate prediction model for user churn prediction. Mini-Batch used in this paper. BP neural network distributed algorithm can speed up the process of training and prediction. At the same time, the characteristics of user correlation can also effectively improve the prediction accuracy. The fast and accurate prediction model proposed in this paper is of great significance to the practical application of the lost user prediction.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN915.0;TP183

【相似文献】