Visual Analytics of Complex Event Sequence Data

事件序列对一段时间内发生的一系列有序事件进行建模。这种序列在广泛的应用中扮演着重要的角色。例如，电子健康记录(EHR)包含在临床过程中记录的患者的带有时间戳的医疗事件(例如，诊断、实验室测试)。作为另一个例子，典型的软件开发过程包括一系列事件，例如提交到存储库。类似地，图形界面的经验评估可以记录用户行为，例如凝视、打字或鼠标点击，这也可以通过事件序列来建模。在这个项目中，我们的目标是开发新的分析和可视化技术，以支持用户深入了解不同应用领域的事件序列数据。我们如何从大量的事件序列中发现事件模式，以便人们能够找到基本规则来帮助解释序列中的因果关系，甚至预测未来的事件？不幸的是，现实世界的事件序列通常规模大、事件类型多样、序列长度不同，并且事件可能以不同的顺序发生和持续不同的持续时间，从而使得事件模式的总结变得困难。我们计划通过创建一种新的视觉分析方法来综合分析复杂事件序列数据，该方法适应和扩展了数据挖掘和机器学习技术，将它们整合到一个交互式视觉分析流水线中。具体地说，我们将首先开发复杂序列数据的预处理和净化技术，以评估事件的重要性，根据底层语义将长序列分割成片段，对不同长度的序列进行比对，并从输入数据中去除噪声。之后，我们将开发一种面向知识的表示学习技术，将各种类型的事件根据它们的语义关系转换为统一的特征表示。利用得到的特征表示，我们将重点介绍用于多尺度序列汇总、事件因果分析、事件预测和异常检测的可视分析技术。我们将把数据预处理、分析和可视化技术集成到针对事件序列数据的可视化分析工具包中。最后，我们将把我们的技术应用到广泛的重要场景中：EHR的新数据分析方法、经验用户研究和软件工程中的事件序列。总体而言，我们期待着基础研究的新结果，这将促进序列数据的可视分析，以及三个应用实例的实际改进。

Event sequences model an ordered series of events that have occurred over a period of time. Such sequences play an important role in a wide range of applications. For instance, electronic health records (EHRs) contain time-stamped medical events (e.g., diagnoses, lab tests) for patients recorded over the course of a clinical process. As another example, a typical software development process includes a series of events such as commits to a repository. Similarly, the empirical evaluation of graphical interfaces may record user behaviors such as gazing, typing, or mouse clicking, which can also be modeled by event sequences. In this project, we aim to develop novel analysis and visualization techniques to support users in obtaining insights into event sequence data in different application domains. How can we discover event patterns from a large collection of event sequences so that people can find the underlying rules to help interpret causality within a sequence and even predict future events? Unfortunately, real-world event sequences are usually large in scale, diverse in event type, vary in sequence length, and events may occur in different orders and last for different durations, thus making the summarization of event patterns difficult. We plan to comprehensively analyze complex event sequence data by creating a novel visual analytics method that adapts and extends data mining and machine learning techniques, integrating them into an interactive visual analysis pipeline.Specifically, we will first develop techniques for preprocessing and cleansing complex sequence data in order to estimate the importance of events, split long sequences into segments based on the underlying semantics, align sequences of different lengths, and remove noise from input data. After that, we will develop a knowledge-oriented representation learning technique that transforms various types of events into a uniform feature representation based on their semantic relations. Utilizing the resulting feature representations, we will then focus on visual analytics techniques for multiscale sequence summarization, event causality analysis, event prediction, and anomaly detection. We will integrate the data preprocessing, analysis, and visualization techniques into a visual analytics toolkit for event sequence data. Finally, we will apply our techniques to a wide range of important scenarios: novel data analysis methods for EHRs, empirical user studies, and event sequences in software engineering. Overall, we expect novel results for basic research that will advance visual analytics of sequence data as well as practical improvements for the three application examples.