基于机器学习的TPTV用户报障预测算法研究

发布时间：2018-03-19 17:20

本文选题：IPTV　切入点：机器学习　出处：《南京邮电大学》2017年硕士论文　论文类型：学位论文

【摘要】：在中国,三网融合正大力推进,交互式网络电视(Internet Protocol Television,IPTV)作为三网融合最合适的着力点,有着十分巨大的潜力,因此对IPTV的研究也成为了当下的热点。然而,运营商传统的运维方法主要是根据用户的投诉来排除设备故障,这种方法时效性差,并且需要大量运维人员,造成人员冗余,已经跟不上时代的发展。为了保证用户收看IPTV的体验,IPTV业务迫切需要一种更合理,更有效的用户报障预测算法作为代替。同时,随着各类计算机性能的迅速提高,机器学习与社会各个领域结合的也越发紧密。本论文从机器学习的角度出发,主要研究了基于机器学习的IPTV故障预测中涉及的一些关键问题,主要的研究内容如下:(1)本论文提出了基于F-Score与互信息的Relief特征选择算法。Relief特征选择算法具有简单明了,运算速度快等优点,并且选择的特征子集具有相当优异的性能,然而它对冗余特征的选择能力较弱。由于Fisher Score对特征的类别也具有很好的区分能力,本论文将Fisher Score加入Relief算法中,以此进一步提高Relief算法的优点,同时为了减少冗余特征,本论文也将互信息与Relief相结合。在多个数据集上的实验表明基于F-Score与互信息的Relief特征选择算法相比原算法的分类准确率得到提高。(2)本论文提出了基于权重限制与F1值的AdaBoost算法。AdaBoost分类算法简单稳定,而且不容易过拟合,针对AdaBoost算法在分类过程中容易对异常点赋予较大权重导致算法失衡和分类错误率不适合用于非均衡数据集的缺陷,本论文对样本的权值做出了限制,并且综合考虑F1值和分类错误率对样本权值的影响,在AdaBoost算法的基础上提出了基于权重限制与F1值的AdaBoost算法,实验表明该算法可以有效提高分类准确率。(3)本论文将基于F-Score与互信息的Relief特征选择算法与基于权重限制与F1值的Ada Boost算法应用于IPTV用户报障预测。本论文对IPTV的各种指标数据进行分析和预处理,然后使用基于F-Score和互信息的Relief算法和基于权重限制与F1值的AdaBoost算法对IPTV数据进行用户报障预测,实验结果表明改进后的算法与原算法相比的预测准确率得到提高。
[Abstract]:In China, tri-network convergence is being vigorously promoted. As the most suitable point of three-network convergence, interactive network television (IPTV) has great potential. Therefore, the research on IPTV has become a hot spot. Operators' traditional operation and maintenance methods are mainly based on customer complaints to troubleshoot equipment, this method is inefficient, and requires a large number of operation and maintenance personnel, resulting in personnel redundancy, In order to ensure users watch the experience of IPTV service, we urgently need a more reasonable and effective algorithm to predict the obstacle of users as a substitute. At the same time, with the rapid improvement of the performance of all kinds of computers, The combination of machine learning and various fields of society is becoming more and more close. From the point of view of machine learning, this paper mainly studies some key problems involved in IPTV fault prediction based on machine learning. The main research contents are as follows: (1) in this paper, a feature selection algorithm for Relief based on F-Score and mutual information. Relief feature selection algorithm has the advantages of simplicity, fast operation and so on, and the selected feature subset has excellent performance. However, its ability to select redundant features is weak. Because Fisher Score also has a good ability to distinguish feature categories, this paper adds Fisher Score to Relief algorithm to further improve the advantages of Relief algorithm and reduce redundant features. This paper also combines mutual information with Relief. Experiments on multiple data sets show that the classification accuracy of Relief feature selection algorithm based on F-Score and mutual information is improved compared with the original algorithm. The AdaBoost algorithm of F1 value. AdaBoost classification algorithm is simple and stable. And it is not easy to fit. In order to solve the problem that AdaBoost algorithm is easy to give outliers a large weight in the process of classification, the algorithm is unbalanced and the classification error rate is not suitable for non-equilibrium data sets, so this paper limits the weight of samples. Considering the influence of F1 value and classification error rate on sample weight, a AdaBoost algorithm based on weight restriction and F1 value is proposed on the basis of AdaBoost algorithm. Experiments show that this algorithm can effectively improve the classification accuracy.) in this paper, the Relief feature selection algorithm based on F-Score and mutual information and the Ada Boost algorithm based on weight limit and F1 value are applied to IPTV user barrier prediction. Analysis and preprocessing of various indicator data, Then the Relief algorithm based on F-Score and mutual information and the AdaBoost algorithm based on weight limit and F1 value are used to predict the IPTV data. The experimental results show that the prediction accuracy of the improved algorithm is higher than that of the original algorithm.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN949.292;TP181

【参考文献】