MOOC用户流失率的统计建模和预测技术的研究

发布时间：2018-05-05 01:02

本文选题：MOOC + 用户流失率　；参考：《北京邮电大学》2016年硕士论文

【摘要】：随着互联网的发展,在线教育蓬勃发展,形式越来越多。MOOC作为在线教育的一种形式,以其学习成本低等优点被越来越多的人接受。但是MOOC面临着一个十分重要的问题:用户流失率过高。这对MOOC的发展十分不利。本文使用了 MOOC网站提供39门课程的用户后台操作数据,对MOOC用户的行为模式进行分析。在此基础上对用户流失率进行建模,实现对MOOC用户是否丢弃课程进行预测。本论文使用的数据集已标注好对应的预测结果(丢弃和不丢弃)。所以对用户流失率建模是一个有监督的二分类问题,可以使用机器学习分类理论来对用户流失率进行建模和预测。论文在分析数据集的基础上,共提出了三类特征,其中有描述性质的特征,也有统计性质的特征,这些特征集合可以从多个角度反映MOOC用户的行为习惯,最大限度地保留了原数据集的信息。接着,论文使用了逻辑回归、支持向量机、随机森林、梯度提升树和AdaBoost分类模型分别对用户流失率进行建模。结合数据集,论证了这五种分类模型的预测性能,同时也指出了分类模型存在的不足:单一的分类模型只能反映数据集一个维度的信息。针对各分类模型独自建模所存在的无法全面反映数据集的不足,论文首先提出了一种多分类器加权联合考虑的思想,并在此基础上给出一种基于分类器权重的改进模型。多分类器加权模型是一种多分类器组合模型,通过加权的方式来组合分类模型,以解决单一分类器不能充分反映数据集的情况。接下来,论文又讨论了一种基于阈值的多分类器与模型,获得一个预测度很高的可信集合。最后,融合多分类器加权模型和可信集合概念,给出一种带有可信集合的加权改算法,进一步提高了流失率模型的预测性能。
[Abstract]:With the development of the Internet, online education is booming, more and more forms of.MOOC as a form of online education are accepted by more and more people with the advantages of low cost of learning. But MOOC is facing a very important problem: the rate of user loss is too high. This is very bad for the development of MOOC. This article uses the MOOC website to provide 3 The users of the 9 courses operate data in the background to analyze the behavior patterns of MOOC users. On this basis, the user churn rate is modeled to predict whether the MOOC users are discarded. The data set used in this paper has been marked with the corresponding prediction results (discarding and discarding). The two classification problem of supervision can be used to model and predict the loss rate of users by machine learning classification theory. On the basis of the analysis of data sets, a total of three types of features are presented, including the characteristics of the description and the characteristics of statistical properties. These feature sets can reflect the behavior habits of MOOC users from multiple angles. Then, the paper uses logical regression, support vector machines, random forests, gradient lifting trees and AdaBoost classification models to model the loss rate of users respectively. The prediction performance of the five classification models is demonstrated by combining data sets, and the shortcomings of the classification models are also pointed out: single The classification model can only reflect the information of one dimension of the data set. In view of the inadequacy of the data sets that can not be fully reflected in the modeling of each classification model, a new idea of weighted joint consideration of multiple classifiers is proposed in this paper, and an improved model based on the weight of classifier is given on this basis. A multi classifier combination model is used to combine the classification model by weighted method so as to solve the case that single classifier can not fully reflect the data set. Then, a multi classifier and model based on threshold is discussed, and a trusted set with high predictability is obtained. Finally, the multiple classifier weighted model and the trusted set are fused. In addition, a weighted modification algorithm with credible sets is proposed to further improve the prediction performance of the churn rate model.

【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：G434;TP181

【相似文献】