移动应用功能描述的评估研究

发布时间：2018-06-16 04:45

本文选题：移动应用 + 功能描述　；参考：《大连理工大学》2016年硕士论文

【摘要】：随着Android操作系统的快速发展,Android应用更新的速度也越来越快,下载量也飞速增长,Google Play商店中移动应用数目已超过200万,为了便于用户了解应用情况,Google Play商店为每个应用提供了一个介绍页面,通过该页面,用户可以了解到该应用的详细信息,包括名称、开发者、截图、功能描述、评论等。为了吸引下载量,很多应用商店会提供一些关于如何书写好的应用功能描述的指导和建议,但是这些建议大多比较抽象或概括,而且也很难定义功能描述的质量好与不好。同时,目前暂没有相关工具自动化地对功能描述的质量进行评估并给出建议。为了评估移动应用功能描述的质量,本文使用了数据驱动的方式构建特征,进而训练评估模型。具体做法是从Google Play中确定了音像(Music Audio)、新闻(News Magazines)、摄影(Photography)、旅行(Travel Local)、天气(Weather)5个类别,从中各选择100个Android应用(Android Applications,以下简称Apps)的功能描述作为样本,然后邀请30个志愿者对样本中功能描述的质量进行评分并说明原因,进而通过对原因的过滤分析,从中构建特征,再结合志愿者的评分训练机器学习模型。通过对志愿者评分原因的分析,我们构建了共计16个特征,结合我们的经验和数据特点以及相关工作确定了各个特征的计算方式,得到的特征值作为输入值；将志愿者对功能描述的评分映射到不同的质量水平(好-Good,中-Neutral,差-Bad),并以此作为输出值,选取了支持向量机(SVM)、决策树(Decision Tree)、随机森林(Random Forest)和逻辑回归(Logistics)模型对所有数据训练测试,进而分析各样本数据在不同模型上的表现。最终音像(Music Audio)类别的样本数据在SVM模型上取得了58%的分类准确率。另外,由于训练模型使用的特征均是我们自己构建,我们对构建的所有特征进行了重要性分析,对比LibSVM模型的特征选择工具和Weka中C4.5决策树模型得到的结果,我们发现在特征重要性排序方面,无监督的学习方式和有监督的学习方式得到的结果基本一致,该结果中有几个特征对于所有样本数据都比较重要：功能描述长度(单词数)、每个单词难易度(每个单词字符个数)、句子长度以及功能描述中应用特征的描述与全部描述的比例,我们希望这个结果对移动应用的开发者有一定启示,即准备移动应用的功能描述时,应注意描述文本的长度、是否易懂以及移动应用的特征描述所占总文本的比例。
[Abstract]:With the rapid development of the Android operating system, the number of mobile apps in the Android play store has exceeded 2 million. The Google play Store provides an introduction page for each application, through which the user can learn the details of the application, including name, developer, screenshot, function description, comment, etc. In order to attract downloads, many app stores provide guidance and advice on how to write a good description of the application's functions, but most of these suggestions are abstract or general, and it is difficult to define whether the quality of the description is good or bad. At present, there are no tools to automatically evaluate the quality of function description and give suggestions. In order to evaluate the quality of functional description of mobile applications, this paper uses a data-driven approach to construct features and then train the evaluation model. This is done by identifying from Google play the five categories of AudioMusic Audio, News Magazinesh, Photography, Travel Local Travel, Weather Weather, and selecting 100 Android applications (hereinafter referred to as Apps) as a sample of their functional descriptions. Then 30 volunteers were invited to score the quality of the function description in the sample and explain the reason. Then through the filtering analysis of the reason the characteristics were constructed and the machine learning model was combined with the evaluation training of the volunteer. Based on the analysis of the cause of volunteer scoring, we constructed a total of 16 features, combined with our experience and data characteristics and related work to determine the calculation method of each feature, the obtained feature value as input value; The scores of function description of volunteers were mapped to different quality levels (good, medium neutral, poor Badn), and as output values, support vector machine (SVM), decision tree (decision tree), random forest random (Random Forest) and logical regression logistic models were selected to test all data training. Then the performance of each sample data in different models is analyzed. Finally, the classification accuracy of audio-video audio-audio category is 58%. In addition, because the features used in the training model are all constructed by ourselves, we analyze the importance of all the features constructed, and compare the feature selection tool of LibSVM model with the results obtained from the C4.5 decision tree model in Weka. We find that the results of unsupervised learning and supervised learning are basically the same as those of supervised learning. There are several features in the result that are important for all sample data: length of function description (number of words), ease of each word (number of characters per word), length of sentence and description of features used in functional description The ratio of the full description, We hope that this result will enlighten the developers of mobile applications, that is, when preparing the functional description of mobile applications, we should pay attention to the length of the description text, whether it is easy to understand and the proportion of the feature description of the mobile application to the total text.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP316;TP181

【相似文献】