Visual analysis of volunteered geographic information for interactive situation modeling and real-time event assessment

在该项目的第一阶段，我们提出了一种新颖的可视化分析方法，使分析师能够根据社交媒体数据的VGI特征获得整体情况评估。特别是，我们研究了如何将可视化交互方案与自动化机器学习工具深度集成，以创建、管理和调整可扩展的ad-hoc回归模型，用于时间关键型决策。这些努力的初步结果，在工作系统原型的形式，已与领域专家在实际的应用场景，以及在实验室条件下进行了评估。从领域专家的反馈，我们得出结论，现有的方法需要在三个主要方向进行扩展。首先，现实世界条件下的情况分析主要是一项协作工作。用于信息收集的模型和查询由各种角色的人员建立、使用和修改，但在工作量方面具有有限的能力。此外，必须以对所有利益攸关方都有意义的方式收集和提出成果。因此，用于特设模型构建和信息收集的高级可视界面应允许多个协作者同时透明地使用，并提供支持不同层次分析深度和决策的定制视角。其次，虽然我们的第一个项目阶段以孤立的方式考虑了新的VGI数据渠道，但如果它支持与其他更传统的信息来源（如新闻媒体）的直接关联和比较，则分析可能会更有价值。语言建模和语言降维的最新成果使得在实时数据流中有效地连接相似或相关的文本信息成为可能。结合数据可视化和交互式推理，这些技术可用于交叉检查来自多个来源的文本内容的重要性和有效性，以及追溯单个信息的来源。第三，在重大事件发生之前、期间和之后，社交媒体渠道中大量出现不准确、错误甚至是欺骗性的信息，这是一个突出的现象，也是形势分析的主要挑战。由于这种错误信息的规模空前--有时只是被毫无戒心的用户无意中分享，有时是由专业团体或通过自动化内容生成软件故意在整个活动中传播--因此需要一些工具来帮助识别正在发生的异常，并评估其空间、时间和拓扑程度。在这种情况下，可视化分析技术允许对数据的VGI特征（例如消息和用户的已知地理位置）进行大规模集成分析，并将它们与条目的文本内容、附加媒体和元数据联合收割机组合，从而提供了一个有希望的杠杆点。

In the first phase of this project, we proposed a novel visual analytics approach that enables analysts to acquire holistic situation assessment based on VGI features of social media data. Particularly, we investigated how the deep integration of visual interactive schemes with automated machine learning tools could be used to create, manage, and adapt scalable ad-hoc regression models for time-critical decision-making. The initial results of these efforts, in the form of working system prototypes, have been evaluated with domain experts in actual application scenarios as well as under lab conditions. From the domain experts' feedback, we conclude that the existing methodology needs to be extended in three major directions. First, situation analysis under real-world conditions is mostly a collaborative effort. Models and queries for information gathering are established, used, and modified by a range of people in various roles but with limited capacities in terms of work load. In addition, results have to be collected and presented in a way that is meaningful to all stakeholders. Accordingly, advanced visual interfaces for ad-hoc model building and information gathering should allow simultaneous, transparent use by multiple collaborators and provide tailored perspectives that support different levels of analytical depth and decision-making. Second, while our first project phase considered novel VGI data channels in an isolated fashion, the analysis can be much more valuable if it supports direct correlation and comparison with other, more traditional sources of information, such as news media. Recent results in language modeling and language dimensionality reduction made it possible to efficiently connect similar or related textual information in real-time data streams. Combined with data visualization and interactive reasoning, these techniques can be used to cross-examine the significance and validity of textual content from multiple sources as well as to trace back the origins of individual pieces of information. Third, the massive occurrence of inaccurate, erroneous, or even deceitful information in social media channels before, during, and after critical events emerged as a salient phenomenon and major challenge for situation analysis. Due to the unprecedented scales of such misinformation - sometimes just inadvertently shared by unsuspecting users, sometimes deliberately distributed in entire campaigns by professional groups or through automated content generation software - there is a need for tools that help to identify ongoing anomalies as well as to assess their spatial, temporal, and topological extent. In this case, visual analytics techniques that allow a large-scale integrated analysis of VGI features of the data, such as known geolocations of messages and users, and combine them with the textual content, attached media, and meta-data of the entries provide a promising leverage point.