基于多信息融合表示学习的关联用户挖掘算法

Associated Users Mining Algorithm Based on Multi-information Fusion Representation Learning

随着互联网技术的迅速发展和普及,越来越多的用户开始通过社会网络进行各种信息的分享与交流。网络中同一用户可能申请多个不同账号进行信息发布,这些账号构成了网络中的关联用户。准确、有效地挖掘社会网络中的关联用户能够抑制网络中的虚假信息和不法行为,从而保证网络环境的安全性和公平性。现有的关联用户挖掘方法仅考虑了用户属性或用户关系信息,未对网络中含有的多类信息进行有效融合以及综合考虑。此外,大多数方法借鉴其他领域的方法进行研究,如去匿名化问题,这些方法不能准确解决关联用户挖掘问题。为此,文中针对网络关联用户挖掘问题,提出了基于多信息融合表示学习的关联用户挖掘算法(Associated Users Mining Algorithm based on Multi-information fusion Representation Learning,AUMA-MRL)。该算法使用网络表示学习的思想对网络中多种不同维度的信息(如用户属性、网络拓扑结构等)进行学习,并将学习得到的表示进行有效融合,从而得到多信息融合的节点嵌入。这些嵌入可以准确表征网络中的多类信息,基于习得的节点嵌入构造相似性向量,从而对网络中的关联用户进行挖掘。文中基于3个真实网络数据对所提算法进行验证,实验网络数据包括蛋白质网络PPI以及社交网络Flickr和Facebook,使用关联用户挖掘结果的精度和召回率作为性能评价指标对所提算法进行有效性验证。结果表明,与现有经典算法相比,所提算法的召回率平均提高了17.5%,能够对网络中的关联用户进行有效挖掘。

With rapid development and popularization of Internet technologies,more and more users have begun to share and exchange various information through social networks.The same user in the network may apply for multiple diffe- rent accounts to distribute information,and these accounts constitute the associated users in the network.Effectively mining associated users in social networks can suppress false information and illegal behaviors in the network,and thus ensure the security and fairness of the network environment.Existing associated user mining methods only consider user attributes or user relationship information without merging multiple types of information contained in the network comprehensively.In addition,most methods draw lessons from the methods in other fields,such as de-anonymization,and they can’t accurately solve the problem of associated user mining.In light of this,this paper proposed an associated user mining algorithm based on multi-information fusion representation learning(AUMA-MRL).In this algorithm,the idea of network representation learning is utilized to learn various dimensional information in the networks,such as user attributes,network topology,etc.Then the learned multi-information is effectively fused to obtain multi-information node embedding,which can accurately characterize multiple types of information in networks,and mine associated users in networks through similarity vectors between node embedding.The proposed algorithm was validated on three real networks namely protein network PPI and social network Flickr,Facebook.In the experiment,the accuracy and recall rate is selected as the performance evaluation indexes.The results show that the recall rate of proposed algorithm is increased by 17.5% on average compared with the existing classical algorithms,and it can effectively mine associated users in networks.