基于小数据的在线用户兴趣长程演化研究  

Study of The Long-Range Evolution of Online Human-Interest Based on Small Data

作　　者：李勇[1,2] 孟小峰[1] 刘继[3] 王常青[4] 

机构地区：[1]中国人民大学信息学院,北京100872 [2]西北师范大学计算机科学与工程学院,兰州730070 [3]新疆财经大学统计与信息学院,乌鲁木齐830012 [4]中国互联网络信息中心互联网基础技术开放实验室,北京100190

出　　处：《计算机研究与发展》2015年第4期779-788,共10页Journal of Computer Research and Development

基　　金：国家自然科学基金项目(61379050,91224008,71261025);国家“八六三”高技术研究发展计划基金项目(2013AA013204);高等学校博士学科点专项科研基金项目(20130004130001);中国人民大学科学研究基金项目(11XNL010)

摘　　要：网络大数据中与Web用户行为相关的数据,例如在线点击数据和通讯记录等,为人们深度挖掘和定量分析人类兴趣动力学带来了机遇,这些在线行为数据被称为大数据时代的"小数据",有助于揭示许多复杂的人类社会与经济现象.Web用户行为建模时常见的前提假设就是人的行为符合Markov过程,用户下一行为仅依赖于当前行为,与过去的历史行为无关.然而,在线用户行为是一个复杂过程,常常依赖于人的兴趣,对于人类兴趣动力学的本质规律目前知之甚少.利用中国互联网络信息中心提供的30 000多名在线用户行为记录数据,基于块熵理论对在线用户行为进行分类研究,通过信息论分析方法,结合熵增曲线的离散导数和积分理论,分析在线用户点击行为的随机性和记忆性特征.研究表明,与常见的假设不同,Web用户的行为并不是一个简单的Markov过程,而是一个符合幂率的非周期无限长程记忆过程;进一步还发现,用户在线连续点击7个兴趣点,其行为的平均预测增益就可达到95.3%以上,可为大数据时代在线用户兴趣精准预测提供理论指导.The availability of network big data ,such as those from online human surfing log ,e‐commerce and communication log , makes it possible to probe into and quantify the dynamics of human‐interest .T hese online behavioral data is called“small data” in the era of big data ,w hich can help explaining many complex socio‐economic phenomena .A fundamental assumption of Web user behavioral modeling is that the user＆#39;s behavior is consistent with the Markov process and the user＆#39;s next behavior only depends on his current behavior regardless of the historical behaviors of the past . However ,Web user＆#39;s behavior is a complex process and often driven by human interests .We know little about regular pattern of human‐interest . In this paper , using more than 30 000 online users behavioral log dataset from CNNIC ,we explore the use of block entropy as a dynamics classifier for human‐interest behaviors . We synthesize several entropy‐based approaches to apply information theoretic measures of randomness and memory to the stochastic and deterministic processes of human‐interests by using discrete derivatives and integrals of the entropy growth curve .Our results are , however preliminary ,that the Web user＆#39;s behavior is not a Markov process ,but a aperiodic infinitary long‐range memory power‐law process .Further analysis finds that the predictability gain can exceed 95 .3 percent w hen users click 7 consecutive points online ,which can provide theoretical guidance for accurate prediction of online user＆#39;s interests in the era of big data .

关 键 词：小数据 块熵 超熵 兴趣演化 预测增益 

分 类 号：TP393.4[自动化与计算机技术—计算机应用技术] TP202.4[自动化与计算机技术—计算机科学与技术]