Data Visualization

考虑以下任务：检测疗养院居民跌倒的风险和缓解措施的处方；监控在线青少年活动以识别潜在的心理健康障碍并建议触发早期治疗的程序；或者确定城市生活质量差异的根源并提出可能的解决方案以等同差异。这些应用程序的关键概念是预测事件，了解预测的原因，并可能制定缓解措施。尽管预测无处不在，使许多应用场景受益，从疾病预测到抵押贷款申请分析，但直到最近，识别偏差和理解结果原因的需求才成为一个令人担忧的问题，政府法规要求解释自动决策。机器学习和视觉分析技术已经被提出来解决这个问题。然而，现有的解决方案孤立地考虑这一过程的不同方面，例如偏见和解释，而很少支持基于预测制定行动的概念。拟议的研究将开发新的视觉分析方法，旨在设计预测性和规范性的工具和技术，以支持信心分类、理解分类和制定缓解措施的整个流程。适当地处理这三个支柱会带来几个挑战，特别是视觉界面的设计，以促进用户参与，通过纳入背景知识使分析过程更加可靠。这项拟议的研究将探索新的视觉隐喻和技术，通过(I)帮助标签过程纠正问题并识别数据偏差，(Ii)帮助理解复杂的模型内部工作和审计结果，以及(Iii)开发说明性视觉分析的新概念，以视觉方式推荐通过预测检测到的问题的潜在解决方案，从而将用户置于构建高质量分类模型的积极参与者的位置。通过产业界和学术界的合作，这项研究的结果将被转化为新的应用，以造福于不同的领域，包括健康监测、城市数据分析等，帮助将视觉分析推广到其他知识领域，并利用Dalhousie在使用、分析、和数据的解释。

Consider the tasks of detecting risks of residents' falls in nursing homes and the prescription of mitigation actions; monitoring online youth activities to identify potential mental health disorders and suggest procedures to trigger early treatments; or identifying the roots for the differences in cities' quality of life and indicate possible solutions to equate the discrepancies.The key concept underlying these applications is predicting an event, understanding the reasons for the prediction, and potentially prescribing mitigation actions. Although prediction is pervasive, benefiting many application scenarios, from disease prognosis to mortgage application analysis, only recently the need for identifying biases and understanding the reasons for an outcome became a concern, with governmental regulations requiring explanations about automated decisions.Machine learning and visual analytics techniques have been proposed to address this issue. However, existing solutions consider the different aspects of this process, e.g., bias and explanation, in isolation while rarely support the concept of prescribing actions based on a prediction. The proposed research will develop new visual analytics approaches aiming to devise predictive and prescriptive tools and techniques to support the whole pipeline of classifying with confidence, understanding a classification, and prescribing mitigation actions.Appropriately addressing these three pillars imposes several challenges, especially the design of visual interfaces to facilitate user involvement, making the analytical process more reliable by incorporating background knowledge. The proposed research will investigate new visual metaphors and techniques to place users as active players in building high-quality classification models by (i) aiding the labeling process to correct problems and identifying data biases, (ii) helping in understanding complex model inner workings and auditing results, and (iii) developing the novel concept of prescriptive visual analytics, visually recommending potential solutions to problems detected through predictions.With industrial and academic collaborations, the results of this research will be translated into novel applications to benefit different domains, including health monitoring, urban data analysis, among others, helping to promote visual analytics into other knowledge domains and leveraging Dalhousie's expertise in the use, analysis, and interpretation of data.