Temporal Causal Pattern Mining From Heterogeneous Event Sequences for Predictive Analytics

互联网不断产生大量的数据流，并形成事件序列，传递有关个人概况、行为、社区、观点、影响、意图和趋势的知识。同样，越来越多的物联网（IoT）传感器产生了大量数据块的连续流，这些数据块通常带有时间戳，也可能带有地理戳，并形成复杂的异构多维事件序列，这些事件序列还传达了大量关于个人身体活动、移动模式、环境变化和医疗信息的知识。从这些复杂的序列数据中提取有趣的模式用于预测分析是一个重要而具有挑战性的问题。从大量复杂的异构序列数据中挖掘时间因果模式进行预测分析是该计划的主要目的。所处理的数据类型是多源的，其特点是异质性、非常大的体积、显著的噪声和缺失值、其属性之间的高维和强相互关系。本研究计划的长期目标是通过挖掘大量序列数据，发展理论，以更好地理解和预测正常和异常行为。通过这个过程，我们的目标也是发展一般理论，可以指导设计有效的序列分析方法，用于现实世界的应用。该计划将通过完成一些相互关联的项目来实施。1)开发一个新的数学框架，用于从事件序列中发现因果关系。将研究深度学习模型，如深度卷积神经网络（CNN）和长短期记忆（LSTM）模型与概率因果关系模型相结合，以发现具有因果关系的统计显著模式。2)开发高效的事件序列异质性挖掘算法。我们将通过将数据投射到新的空间来创建新的潜在数据表示，作为抽象，这将使学习更容易。这里开发的算法将能够提取复杂的因果模式和各种事件序列之间的相互作用。3)开发有效的算法来预测社交网络中的暴力行为。这个项目是我们在这个研究项目中开发的理论的一个重要应用，它可以解决一个实际的现实世界问题，并有可能为社交网络用户产生有益的社会影响。它也将作为这个研究项目的试验台。

The Internet continuously produces large flows of data and forms events sequences that convey knowledge about individuals' profiles, behaviors, communities, opinions, influences, intentions, and trends. Similarly, the increasing number of Internet of Things (IoT) sensors generate a continuous flow of large amounts of data chunks that are usually time-stamped and possibly geo-stamped, and form complex heterogeneous multidimensional event sequences conveying also a great deal of knowledge about individuals' physical activities, mobility patterns, environmental changes, and medical information. Extracting interesting patterns from these complex sequence data for predictive analytics is an important and challenging problem.Mining temporal causal patterns from massive and complex heterogeneous sequence data for predictive analytics is the main purpose of this program. The data types addressed are multisource, characterized by heterogeneity, very large volume, significant noise and missing values, high-dimensional and strong interrelations between their attributes. The long-term goal of this research program is to develop theories to better understand and predict normal and abnormal behaviors through mining massive sequence data. Through this process, our goal is also the development of general theories that can guide the design of efficient sequence analysis methods for real-world applications.This program will be carried out by accomplishing a number of interrelated projects. 1) Develop a novel mathematical framework for causality discovery from event sequences. Deep learning models such as deep Convolutional Neural Networks (CNN) and Long-Short Term Memory (LSTM) models combined with probabilistic causality models will be investigated to discover statistically significant patterns with causal relationships. 2) Develop efficient algorithms for mining heterogeneity of event sequences. We will create new latent representations of data that serve as abstractions by projecting data into new spaces, which will make learning easier. The algorithms developed here will be able to extract complex causal patterns and interactions between various event sequences. 3) Develop efficient algorithms for predicting violent behaviors in social networks. This project is an important application of the theory we developed in this research program to solve a practical real world problem, and would have the potential to generate a beneficial social impact for social network users. It will also serve as a testbed of this research program.