[论文解读] Cluster Analysis for a Scale-Free Folksodriven Structure Network

Massimiliano Dal Mas|arXiv (Cornell University)|Dec 19, 2011

Complex Network Analysis Techniques参考文献 23被引用 10

一句话总结

本文提出一种动态加权的协作标签网络——'Folksodriven'——其中标签通过用户、网站及时间维度上的共现频率相互关联。通过聚类分析与网络度量方法，研究发现该结构具有无标度特性、高度连通，且能有效支持偶然的内容发现，其聚类系数与路径长度表现优于随机网络。

ABSTRACT

Folksonomy is said to provide a democratic tagging system that reflects the opinions of the general public, but it is not a classification system and it is hard to make sense of. It would be necessary to share a representation of contexts by all the users to develop a social and collaborative matching. The solution could be to help the users to choose proper tags thanks to a dynamical driven system of folksonomy that could evolve during the time. This paper uses a cluster analysis to measure a new concept of a structure called "Folksodriven", which consists of tags, source and time. Many approaches include in their goals the use of folksonomy that could evolve during time to evaluate characteristics. This paper describes an alternative where the goal is to develop a weighted network of tags where link strengths are based on the frequencies of tag co-occurrence, and studied the weight distributions and connectivity correlations among nodes in this network. The paper proposes and analyzes the network structure of the Folksodriven tags thought as folksonomy tags suggestions for the user on a dataset built on chosen websites. It is observed that the hypergraphs of the Folksodriven are highly connected and that the relative path lengths are relatively low, facilitating thus the serendipitous discovery of interesting contents for the users. Then its characteristics, Clustering Coefficient, is compared with random networks. The goal of this paper is a useful analysis of the use of folksonomies on some well known and extensive web sites with real user involvement. The advantages of the new tagging method using folksonomy are on a new interesting method to be employed by a knowledge management system. *** This paper has been accepted to the International Conference on Social Computing and its Applications (SCA 2011) - Sydney Australia, 12-14 December 2011 ***

研究动机与目标

• 开发一种随时间演化的动态用户驱动标签系统，基于大型网站的真实用户数据。
• 将协作标签系统建模为加权网络，其中边的权重反映标签共现频率。
• 分析聚类系数、路径长度与连通性等结构特性，评估网络在内容发现中的效率。
• 将协作标签网络与随机网络进行对比，验证其鲁棒性与可扩展性。
• 为知识管理系统提供一个框架，支持协作式、上下文感知的标签推荐。

提出的方法

• 构建一个加权网络，其中节点代表标签，边代表用户生成元数据中标签的共现频率。
• 利用时间与来源（网站）的元数据丰富网络上下文，形成 'Folksodriven' 结构。
• 应用聚类分析以度量网络凝聚力，并识别密集的标签社区。
• 计算聚类系数与平均路径长度，以评估网络拓扑结构与效率。
• 将实证网络与随机网络模型进行对比，评估其无标度与小世界特性。
• 使用知名网站的真实世界数据集，验证模型在真实用户行为下的有效性。

实验结果

研究问题

• RQ1 当引入时间、来源与用户上下文后，协作标签网络的结构如何演化？
• RQ2 基于共现的标签网络中，边权重的分布特征是什么？是否表现出无标度特性？
• RQ3 Folksodriven 网络的聚类系数与随机网络相比如何？
• RQ4 Folksodriven 网络的平均路径长度是多少？是否支持高效的内容偶然发现？
• RQ5 该网络结构能否支持高效、动态的标签推荐系统？

主要发现

• Folksodriven 网络表现出无标度结构，节点度数呈幂律分布，表明存在高度连接的 '枢纽' 标签。
• 网络具有较高的聚类系数，显著高于随机网络，表明存在强局部连通性与社区形成。
• 平均路径长度相对较短，支持高效导航与相关资源的偶然发现。
• 网络高度连通，标签形成密集的超图结构，支持稳健的信息检索与推荐。
• 标签共现频率构成一个稳定且加权的网络，随用户活动动态演化，支持实时推荐。
• 结构特性证实该网络适用于协作式知识管理系统与智能标签系统。