Time-sensitive non-parametric Bayesian approaches for events modeling, recognition and prediction

事件的建模、识别和预测是多个学科的核心问题。 这些都是具有挑战性的任务，具有重要的经济、社会、技术和安全优势。过去已经提出了几种机器学习和数据挖掘技术来从数据中提取有用的知识和隐藏模式，以识别事件并预测未来的事件（例如机会，威胁等）。这些技术的广泛使用，以解决不同的任务，使从业者能够弄清楚，事件通常是复杂的性质（即联合收割机数据来自不同的来源），传统的事件建模，识别和预测方法，一般是两个假设，往往是错误的，在现实生活中的应用。第一，事件类别的数量是有限的和已知的。第二，表示事件的特征向量是独立的，这在大多数情况下是不正确的。 一般来说，事件顺序到达，并且具有一定的时间依赖性，因为它们依赖于随时间变化的数据。 该研究计划的目标是探索和开发大规模的统计学习方法，通过分析时变数据来建模，识别和预测事件。 特别是，我将专注于非参数贝叶斯模型。 这些模型是灵活的，消除了对固定数量的事件类的假设的需要，但不幸的是，是可交换的（即旧事件与新事件具有同等重要性），并且不捕获随时间变化的特征。然后，我建议将时间的概念引入到非参数贝叶斯模型中，以保持手头事件的最新表示。虽然在大多数现有的统计框架中，对特征向量进行了特定的分布假设，但我建议考虑分布族，以增加更多的灵活性和泛化能力。这些扩展需要开发新的学习方法来估计所开发的模型的参数和设计有效的特征选择技术。 拟议的研究计划是非常有前途的，基于强大的数学和统计公式。它将促进知识的发展，提高技术水平，并有可能实现许多应用。例子包括社交网络的多媒体内容（文本，图像，视频）的建模，犯罪威胁预测，视频中的异常事件识别等，我提出的研究结果可以应用于不同的其他领域和学科（例如，生物学，社会科学等）。

The modeling, recognition and prediction of events are central problems of several disciplines. These are challenging tasks that have important economic, social, technological and security advantages. Several machine learning and data mining techniques have been proposed in the past to extract useful knowledge and hidden patterns from data to recognize events and to predict future ones (e.g. opportunities, threats, etc.). The extensive use of these techniques, to tackle different tasks, has allowed practitioners to figure out that events are generally complex in nature (i.e. combine data from different sources) and that traditional events modeling, recognition and prediction approaches make generally two assumptions that are often false in real-life applications. First, that the number of classes of events is finite and known. Second, that the feature vectors representing the events are independent which is not true in the majority of the cases. Generally, events arrive sequentially and have certain dependency in time since they depend on time-varying data. The goal of this research program is to explore and develop large-scale statistical learning approaches to model, recognize and predict events by analyzing time-varying data. In particular, I will focus on nonparametric Bayesian models. These models are flexible, eliminate the need for an assumption of a fixed number of classes of events, but are unfortunately exchangeable (i.e. old events have equal importance to new ones) and do not capture time-varying features. I propose then to introduce the notion of time into nonparametric Bayesian models to maintain an up-to-date representation of the events at hand. While in most existing statistical frameworks specific distributional assumptions are made for the feature vectors, I propose to consider families of distributions to add more flexibility and generalization capabilities. These extensions require the development of new learning approaches to estimate the parameters of the developed models and the design of efficient feature selection techniques. The proposed research program is very promising, based on strong mathematical and statistical formulations. It will advance knowledge and improve the state of the art. Many applications are possible. Examples include the modeling of the multimedia content (text, image, video) of social networks, criminal threats prediction, abnormal events recognition in videos, etc. The results of the research that I propose can be applied for different other fields and disciplines (e.g. biology, social science, etc.).