BIGDATA: Small DA Social Behavior Driven Modeling and Optimization of Information

描述（由申请人提供）：问题：大规模的社交媒体和社交互动数据，如推文，博客，论坛，变得越来越可用。信息在社交网络和社交媒体上的传播模式通常是隐藏的。对这些历史社会互动的建模，为理解和优化社交网络中的信息传播提供了巨大的潜力。这种模式也有实际的影响，如。促进卫生保健讨论论坛的活动，并加速在科学界传播思想。然而，大多数以前的社会网络分析方法侧重于定性和宏观的解释性分析的网络行为，而不是定量和微观的预测模型。这些模型难以用于后续的信息扩散优化和管理。因此，有一个强大的和预测性的建模框架，利用大规模的历史社会互动数据，可以适应社会互动的复杂性和异质性的需求。目的：本项目的目标是开发一套强大的 基于复杂和噪声交互数据的机器学习方法，用于建模和优化信息扩散过程。它由四个部分组成：（i）开发一个新的概率框架，用于对社交网络中的事件级联进行建模和推理;（ii）开发非参数核方法，以捕获社会交互的复杂性和异质性;（iii）开发有效的在线/批量优化算法，用于从大型数据集估计扩散模型;以及（vi）使用所估计的模型的预测来优化信息传播并促进社会互动。技术创新和优点：我们将在社交网络环境中使用通常用于医疗数据分析的事件历史分析。这为我们提供了一个原则性和全面的框架，以解决我们项目中的所有四个方面。事件历史分析和核方法的结合还揭示了信息扩散建模问题和分组套索统计估计问题之间的联系，使我们能够将最近发展的稀疏恢复理论引入到社会网络问题中，如发现信息扩散渠道，并正式研究这种恢复的条件和统计保证。相关性（参见说明）：我们提出的研究在健康讨论论坛中有广泛的应用;由糖尿病手基金会运营的TuDiabetes将是一个试验平台。该项目有可能提高人们在讨论论坛中的参与度，并为糖尿病患者提供更好的社会福利。本研究也将事件历史分析、核方法、图模型、稀疏性恢复理论等多个研究领域结合起来研究社会网络问题。

DESCRIPTION (provided by applicant): The Problem: Large-scale social media and social interaction data, such as tweets, blogs, discussion forums, are becoming increasing available. The patterns of information diffusion across social networks and social media are generally hidden. Modeling these historical social interactions, promises great potentials for the understanding and optimization of information diffusion in social networks. Such models also have practical impacts such as. promoting activities in health care discussion forums and accelerate the dissemination of ideas in scientific communities. However, most previous approaches for social network analysis focus on qualitative and macroscopic explanatory analysis of the network behavior, rather than quantitative and microscopic predictive models. It is difficult to make use of these models for subsequent optimization and management of information diffusion. Thus there is a great need for a robust and predictive modeling framework leveraging the large-scale historical social interaction data and can adapt to the complexity and heterogeneity of social interactions. , Aim: The goal of this project is to develop a set of robust machine learning methods for modeling and optimizing the information diffusion processes, based on the complex and noisy interaction data. It consists of a pipeline of four components: (i) develop a novel probabilistic framework for modeling and reasoning about cascades of events in social networks; (ii) develop nonparametric kernel methods to capture the complexity and heterogeneity of social interaction; (iii) develop efficient online/batch optimization algorithms fr estimating the diffusion models from large datasets; and (vi) optimize information diffusion and promote social interaction using the predictions of the estimated models. Technical Innovation and Merit: We will make novel use of event history analysis typically used for medical data analysis in the social network context. This provides us a principled and over-arching framework for addressing all four aspects in our project. The combination of event history analysis and kernel methods also reveals the connection between the information diffusion modeling problem and the grouped lasso statistical estimation problem, allowing us to bring in recently developed sparse recovery theory into social network problems such as discovery of information diffusion channels and formally study the conditions and statistical guarantees for such recovery. RELEVANCE (See instructions): Our proposed research has wide-ranging applications in health discussion forum; TuDiabetes which is operated by the Diabetes Hands Foundation will be a testbed. This project has the potential to improve the engagement of people in the discussion forum and foster better social goods for diabetes patients. The proposed research also bring together several research areas, such as event history analysis, kernel methods, graphical models, and sparsity recovery theory, to study social network problems.