De Masi Alexandre, Wac Katarzyna
University of Geneva, Geneva, Switzerland.
University of Copenhagen, Copenhagen, Denmark.
Qual User Exp. 2020;5(1):10. doi: 10.1007/s41233-020-00039-w. Epub 2020 Oct 4.
Progressively, smartphones have become the pocket Swiss army knife for everyone. They support their users needs to accomplish tasks in numerous contexts. However, the applications executing those tasks are regularly not performing as they should, and the user-perceived experience is altered. In this paper, we present our approach to model and predict the Quality of Experience (QoE) of mobile applications used over WiFi or cellular network. We aimed to create predictive QoE models and to derive recommendations for mobile application developers to create QoE aware applications. Previous works on smartphone applications' QoE prediction only focus on qualitative or quantitative data. We collected both qualitative and quantitative data "in the wild" through our living lab. We ran a 4-week-long study with 38 Android phone users. We focused on frequently used and highly interactive applications. The participants rated their mobile applications' expectation and QoE and in various contexts resulting in a total of 6086 ratings. Simultaneously, our smartphone logger (mQoL-Log) collected background information such as network information, user physical activity, battery statistics, and more. We apply various data aggregation approaches and features selection processes to train multiple predictive QoE models. We obtain better model performances using ratings acquired within 14.85 minutes after the application usage. Additionally, we boost our models' performance with the users expectation as a new feature. We create an on-device prediction model with on-smartphone only features. We compare its performance metrics against the previous model. The on-device model performs below the full features models. Surprisingly, among the following top three features: the intended task to accomplish with the app, application's name (e.g., WhatsApp, Spotify), and network Quality of Service (QoS), the user physical activity is the most important feature (e.g., if walking). Finally, we share our recommendations with the application developers, and we discuss the implications of QoE and expectations in mobile application design.
智能手机逐渐成为每个人口袋里的瑞士军刀。它们满足用户在多种场景下完成任务的需求。然而，执行这些任务的应用程序常常表现不佳，从而改变了用户的感知体验。在本文中，我们展示了我们对通过WiFi或蜂窝网络使用的移动应用程序的体验质量（QoE）进行建模和预测的方法。我们旨在创建预测性QoE模型，并为移动应用程序开发者提供建议，以开发具有QoE意识的应用程序。以往关于智能手机应用程序QoE预测的工作仅关注定性或定量数据。我们通过生活实验室在“实际使用场景中”收集了定性和定量数据。我们对38名安卓手机用户进行了为期4周的研究。我们聚焦于常用且高度交互的应用程序。参与者对他们的移动应用程序在各种场景下的期望和QoE进行评分，总共得到6086个评分。同时，我们的智能手机日志记录器（mQoL-Log）收集网络信息、用户身体活动、电池统计数据等背景信息。我们应用各种数据聚合方法和特征选择过程来训练多个预测性QoE模型。我们使用应用程序使用后14.85分钟内获得的评分获得了更好的模型性能。此外，我们将用户期望作为一个新特征来提升模型性能。我们创建了一个仅使用智能手机上的特征的设备端预测模型。我们将其性能指标与之前的模型进行比较。设备端模型的表现低于全特征模型。令人惊讶的是，在以下三个最重要的特征中：使用应用程序要完成的预期任务、应用程序的名称（如WhatsApp、Spotify）和网络服务质量（QoS），用户身体活动是最重要的特征（如步行时）。最后，我们与应用程序开发者分享我们的建议，并讨论QoE和期望在移动应用程序设计中的影响。
