Research On Representation Learning For Structural Relation In Social Networks

Internet applications have been deeply rooted in all aspects of human society.The online socialization has become an indispensable element in most people’s lives.Since a large number of users are flooding into the Internet,the social networking platforms are gradually gathering all kinds of information,such as social relationships,personal behaviors,developing tracks and so on.Different individuals in the Internet are connected to each other based on related association rules,which form the(online)social networks.Thus,structural relationship is regarded as the skeleton underneath the social networks,which reveals the propagation of information in networks.The innovation of information technology has incentivized the development of intelligent social networking services.Recently,basic functions of network services,for example message dissemination or content sharing,cannot satisfy the demands of users in modern social networks.Throughout the research on social networks,traditional methods have been unable to be well-applied to the analysis of complex network structures to some extent.Currently,deep learning techniques are leading the era of artificial intelligence.However,intelligent analysis on structural relationships of social networks still needs to be further explored.Specifically,from the perspective of graph theory,the relationship mining models still have room to be improved.Besides,the cross-modal or multi-modal graph structure relationship mining has been rarely noticed.Therefore,this paper aims at developing deep-learning-based network representation learning technologies to analyze and discover structural relationships in networked systems.Following the trend of current research efforts,the main contributions of this paper can be summarized into three aspects:(1)Research on network representation learning method based on structural similarity in proximity space of networks.In order to fully exploit the properties of context structures in local neighborhood space of networks,this paper proposes a network representation learning method,namely SimWalk,based on Markov Chain walk guided by node-to-node similarities.Inspired by word2vec,one of the deeplearning-based natural language processing technologies,SimWalk measures the node-to-node similarities to infer the probability transition matrix in the local network space which guides the graph structure exploration process based on Markov Chain.Therefore,analogous to the textual semantic structures in natural language,the generation procedure of node sequences can highlight the co-occurrence and correlation of similar nodes.Finally,the deep learning model is applied to encode the context structure attributes into the network representations.Experimental results show that the node representations generated by SimWalk can better preserve the structural relationships in the networks in comparison with the network representation learning baselines;(2)Research on network representation learning method based on deep isomorphic substructure encoding from network topologies.To deal with the complex structure of modern social networks and to learn more comprehensive information of structural relationship from the network,this paper proposes a Deep Isomorphic Substructure Learning(DISL)method,which can perceive structural relationships in networks.According to the substructural partition based on isomorphism,DISL makes the isomorphic units self-embedded into the node-based subgraphs,the potential similar features thereby can be preserved according to the geometric forms.In addition,the SimWalk method is applied to explore the context structure of the local network spaces to capture the co-occurrence and correlation of similar nodes.Finally,DISL uses Convolutional Neural Networks to encode both the geometric morphological features and local context structure features to learn a more comprehensive low-dimensional vector representation.Compared with the state-of-the-art network representation learning baseline methods,Experimental results present that the network representations generated by DISL can better reflect the structural relationships of networks;(3)Research on network representation learning method for cross-modal structural relation inference in text-abundant information networks.Since most existing network representation learning methods have to largely depend on the "observed connections",i.e.,the topological attributes in network structures,some valuable social relationships might not be discovered due to the lack of structural information.Since the task of inferring valid social relations in text-based information networks has drawn great attention recently,this paper proposes a Social Relation Generative Adversarial Networks(SRGAN)method,which aims to effectively utilize text data to infer meaningful social relations in text-based networked systems with incomplete structural features.SRGAN breaks the barrier between graph modal data and text modal data by the adversarial learning process so that the cross-domain knowledge can be mapped between the two modalities.As a consequence,SRGAN can encode text modal data to construct missing structures or predict potential social relationships.Experimental results show that SRGAN can provide more realistic social relationships in rich text information networked systems compared against the state-of-the-art methods.