With lots of public software descriptions emerging in the application market, it is significant to extract common software features from these descriptions and recommend them to new designers. However, existing approaches often recommend features according to their frequencies which reflect designers' preferences. In order to identify those users' favorite features and help design more popular software, this paper proposes to make use of the public data of users' ratings and products' downloads which reflect users' preferences to recommend extracted features. The proposed approach distinguishes users' perspective from designers' perspective and argues that users' perspective is better for recommending features because most products are designed for users and expect to be popular among users. Based on the lasso regression to estimate the relationship between the extracted features and the users' ratings, it first distinguishes the extracted features to identify those recommendable and undesirable features. By treating each download as a support from users to the product featurefeatures, it further mines the feature association rules from users' perspective for recommending features. By taking the public data on the market of SoftPedia.com for evaluation, our empirical studies indicate that: (i) selecting recommendable features by lasso regression is better than that by feature frequencies in terms of F1measure; and (ii) recommending features based on the feature association rules mined from users' perspective is not only feasible but also has competitive performance compared with that based on the rules mined from designs' perspective in terms of F1measure.

中文翻译：

---

向设计师推荐软件功能：从用户的角度

随着应用程序市场中出现了大量的公共软件描述，从这些描述中提取常见的软件功能并将其推荐给新设计人员非常重要。然而，现有方法通常根据反映设计者偏好的频率推荐特征。为了识别这些用户最喜欢的特征并帮助设计更受欢迎的软件，本文提出利用反映用户偏好的用户评分和产品下载的公共数据来推荐提取的特征。所提出的方法将用户的观点与设计者的观点区分开来，并认为用户的观点更适合推荐功能，因为大多数产品都是为用户设计的，并希望在用户中流行。基于套索回归估计提取的特征与用户评分之间的关​​系，首先区分提取的特征以识别那些可推荐和不受欢迎的特征。通过将每次下载视为用户对产品功能特性的支持，进一步从用户的角度挖掘特征关联规则，用于推荐功能。通过SoftPedia.com市场上的公开数据进行评估，我们的实证研究表明：（i）在F1measure方面，通过套索回归选择推荐特征优于通过特征频率选择推荐特征；(ii) 基于从用户的特征中挖掘出的特征关联规则来推荐特征