III-COR+RI: Novel Statistical Models and Algorithms for Network Modeling, Mining and Reverse Engineering

在生物学、社会科学和其他领域出现的许多问题中，经常需要分析由网络相互连接的实体（如分子或个体）群体。本提案旨在开发新的统计形式和计算方法来建模和推断网络实体的语义基础，并研究这些方面如何影响网络拓扑及其在生物和社会学过程中的时间演变。它还将研究一些尚未探索的主题，如网络结构的判别学习，恢复时间进化的网络序列以及相关的理论问题。提出的研究设想有助于解决诸如：1)隐藏的身份/功能归纳，例如，当个体在不同条件下与不同的同伴互动时，他们扮演什么角色？2)结构/组织预测，例如，分子功能的改变是否以及如何导致生物通路的改变？3)系统鲁棒性，例如，网络如何适应外源入侵引起的扰动？这项研究跨越了统计学习、社会/生物科学和数据挖掘。所提议的工作的智力价值在于方法发展的算法和理论的新颖性，以及对特定社会和生物网络的分析以及所提议的方法所能实现的各种其他应用。主要创新点包括：(1)节点函数和网络连接的潜在空间建模的新贝叶斯形式化，它捕获了网络实体的功能/行为上下文；(2)基于指数随机图模型的网络演化时间扩展及推理/学习算法；(3)基于纵向节点属性数据的时间重构网络逆向工程算法；(4)从网络的部分样本中学习超大型网络的新型判别学习算法和相关学习理论。这些方法将应用于安然电子邮件网络，以探索各种商业运作条件下的行为模式，并分析从乳腺癌细胞测量的纵向分子丰度剖面，以推断在致癌或肿瘤抑制环境下网络的（改变）。研究结果有望推动网络分析原理和技术的发展，并实现更广泛的应用。拟议中的研究也有望产生广泛的教育和社会影响。作为一项跨学科的研究工作，该项目将为本科生和研究生提供丰富的多学科教育和研究培训机会。对人类社会网络结构的透彻理解可以对政策制定或技术采用等重要问题产生重大影响。了解细胞网络及其响应外源干预的变化可以帮助推理疾病原因和设计治疗方案。我们的方法和软件交付成果可以潜在地促进此类研究，提高网络数据收集的成本效益，并促进该领域的未来发展。这个项目的更多细节可以在http://www.cs.cmu.edu/~epxing/projects/network.htm上找到

In many problems arising in biology, social sciences and various other fields, it is often necessary to analyze populations of entities (e.g., molecules or individuals) interconnected by a network. This proposal intends to develop new statistical formalisms and computational methodologiesfor modeling and inference the semantic underpinnings ofnetwork entities, and investigate how these aspects influence the network topology and its temporal evolution during biological and sociological processes. It will also study a number of yet unexplored topics such as discriminative learning of network structures,recovering temporally evolving network sequences, and related theoretical issues. The proposed research is envisaged to help address big-picture problems such as: 1) Hidden Identity/Function Induction, e.g., what role(s)do individuals play when they interact with different peers under differentconditions? 2) Structural/Organizational Forecast, e.g., whether and how changes of molecular functions lead to alterations of biological pathways? 3) System Robustness, e.g., how a network adjusts to perturbations caused by exogenous intrusions? This research straddles statistical learning, social/biological sciencesand data mining. The intellectual merit of the proposed work lies in both the algorithmic and theoretical novelties of the methodological developments, and the analysis of specific social and biological networks and various other applications enabled by the proposed methods. The main novelties include: (1) new Bayesian formalisms for latent space modeling of node functions and network linkages, which capture the functional/behavioral context of network entities; (2) novel temporal extensions of exponential random graph model for network evolution, and inference/learning algorithms; (3) algorithms for reverse-engineering temporally rewiring networks from longitudinal node attribute data; and (4) novel discriminative learning algorithms for learning very-large networks from partial samples of the network and relevant learning theory. These methods will be applied tothe ENRON email network to explore the behavioral patterns under various business operation conditions, and to analyze a longitudinal molecular abundance profile measured from breast cancer cells to infer (alterations of) networks under carcinogenic or tumor-suppressing environments. The results are expected to advance the principles and technologies for network analysis, and enable a wide-range of applications of broader interests.The proposed research is also expected to have broad educational and societal impact. As an interdisciplinary research effort, this project will provide rich opportunities for multi-disciplinary educational and research training, at both undergraduate and graduate levels. A thorough understanding of social network structures in human populations can have significant impact on important issues such as policy making or technology adoption. Knowledge of cellular networks and its changes in response to exogenous interventions can help reasoning disease causes and designing therapeutic schemes. Our methodologicaland software deliverables can potentially facilitate such studies, improve thecost-effectiveness of network data collection, and foster future developmentin this area. More details of this project can be found at http://www.cs.cmu.edu/~epxing/projects/network.htm