Given the popularity of Internet and new technology,more and more behavioral data recording human interactions has now become available,and attracted the attention of sociological research.Most of the behavioral journal data are of event-action type and are the same data structure as two-mode networks.Two-mode networks are common in social network analysis fields and there are many methods for analyzing two-mode networks.However,unlike the classical two-mode network that is usually a small dataset and suitable for methods such as matrix decomposition,principal component analysis,and other descriptive analysis methods,the underlying network of behavioral data is rather large in scale,with information about time ordered heterogeneous events.Besides,the network members change dynamically,members may join or leave the network.Traditional analytic methods cannot effectively deal with such data.The analysis of such large-scale behavioral data is a huge challenge for social scientists.Over the past decade, the high dimensional Gaussian graphic model has received a great deal of attention in the research of network structure detection,especially those based on Tibshirani's lasso method of statistical analysis(1996).The success of the lasso based penalized Gaussian graphic model is not only due to its efficiency in high dimensional computation,but also due to its interpretability and ease of extension under further considerations.Hence,the lasso penalized Gaussian graphic model is a rapidly developing field with an overwhelming amount of literature on Biology,Genetics,Neurology,machine learning,etc.However,it hasn't caught the attention from social scientists.This paper presents an overview of the applications of lasso based penalized Gaussian graphic model for the measurement of network structures with observational behavioral data.The author does not focus on the specific solution algorithms and optimization processes,but rather on the potential substantial contributions of the Gaussian graphic model and its extensions to social science research.This paper derives different hypothesis under theoretical concern and demonstrates with real data examples.Finally,it also briefly summarizes the related models and their R packages,with intent to expand the application of the Gaussian graphic models in social science research.%随着互联网及智能设备的普及,越来越多的用户行为轨迹和互动数据的获得成为可能并进入社会学研究者的视野.这类行为或互动事件的数据在数据结构上属于社会网络分析方法中常见的双模网络.但传统的社会网络分析所面对的数据规模较小,研究者一般采用矩阵分解、主成分分析等描述性分析方式来对网络子群进行区分或测量.而在大数据的背景下,参与互动的群体规模巨大、群体成员的构成动态变化、事件具有时序特征、事件发生存在异质性等特征,使得传统的分析方法无法有效应对此类数据. 近十年来,高维高斯图模型在网络关系探测研究中被广泛应用.本文拟对基于罚似然回归的高斯图模型进行综述.罚似然高斯图模型是一个发展迅速的分析工具,本文并不侧重具体的算法和优化过程,而是就罚似然图模型及其扩展模型对社会科学应用研究可能带来的贡献进行梳理.最后,本文亦对涉及的相关模型及其R软件包进行汇总,以期拓展该方法在社会科学领域的应用.
展开▼
