符号网络的局部标注特征与预测方法

PDF下载HTML

Title:

Local labeling features and a prediction method for a signed network

作者:

苏晓萍¹, 宋玉蓉²

1. 南京工业职业技术学院 计算机与软件学院, 江苏 南京 210046;
2. 南京邮电大学 自动化学院, 江苏 南京 210003

Author(s):

SU Xiaoping¹, SONG Yurong²

1. School of Computer and Software Engineering, Nanjing Institute of Industry Technology, Nanjing 210046, China;
2. College of Automation, Nanjing University of Posts and Telecommunications, Nanjing 210003, China

关键词:

符号网络; 符号预测; 低秩; 矩阵分解; 标注偏置; 结构平衡理论; 弱结构平衡理论; 地位理论

Keywords:

signed networks; sign prediction; low rank; matrix factorization; signed bias; structural balance theory; weak structural balance theory; status theory

分类号:

TP399

DOI:

10.11992/tis.201710027

摘要:

当复杂网络的边具有正、负属性时称为符号网络。符号为正表示两用户间具有相互信任（朋友）关系，相反，符号为负表示不信任（敌对）关系。符号网络中的一个重要研究任务是给定部分观测的符号网络，预测未知符号。分析发现，具有弱结构平衡特征的符号网络，其邻接矩阵呈现全局低秩性，在该特征下链路符号预测问题可以近似表达为低秩矩阵分解问题。但基本低秩模型中，相邻节点间符号标注的局部行为特征未得到充分利用，论文提出了一种带偏置的低秩矩阵分解模型，将邻居节点的出边和入边符号特征作为偏置信息引入模型，以提高符号预测的精度。利用真实符号网络数据进行的实验证明，所提模型能够获得较其他基准算法好的预测效果且算法效率高。

Abstract:

A complex network may be considered as a symbol network when links have a positive or negative sign attribute. In signed social networks, positive links represent a trust (friends) relationship between users. In contrast, negative links indicate distrust (hostility). An important task in a signed network is to define a signed network based on partial observation to predict an unknown symbol. Through analysis, we found that for a signed network with weak structural balance, its adjacent matrix has a global low-rank characteristic and the prediction of the link sign can be approximated as a low-rank matrix factorization. However, in a basic low-rank model, it is difficult to sufficiently utilize the local behavior features for labeling the signs of links between the neighboring nodes. Herein, a low-rank matrix factorization model with bias was proposed. In this model, the sign features of the exit and entry links of a neighboring node were introduced to improve the precision of sign prediction. Experiments based on real data revealed that the low-rank model with bias can obtain better prediction results than other benchmark algorithms and that the proposed algorithm performed with a high efficiency.