Research And Implementation Of Group Discovery Based On Activity Semantics

Online social networks have become an important platform for people to communicate with each other.There are some users who take advantage of the anonymity of the Internet to spread illegitimate remarks,organize and plan vicious incidents,which is a great challenge to public security governance.However,when identifying and tracking these special user groups,the following problems remain:first,the current mainstream group discovery methods are oriented towards general groups and lack pertinence in user behavior analysis.And it’s not efficient enough to mine the characteristics of domain users with a single behavior relationship network;Secondly,when using representation learning for efficient social network data analysis,there is still an obvious imbalance in the topological and semantic information;Finally,the semantic shift in social network group discovery may also lead to the fact that the final group is not the target group or just a part of the target group.To solve the above problems,this paper has carried out the following research work:1.Aiming at the vulnerability of single behavior relationship network,this paper models the online social network based on activity semantics,and designs a domain-oriented multi-dimensional activity semantic extraction model.We extract activity semantics from multi-dimensional and multi-level by using ConSERT contrastive learning model,fastText short text emotion classification model and domain features designed,The complex multi-modal online social network is decomposed into multiple single-modal sub-networks,so that the same user group presents different topological associations and attribute characteristics under different activity semantics.It better reflects relationships and states diversity of the real social network,and provides a basis for discovery of special groups.2.Aiming at the imbalance between topology and semantics in the embedded representation of social networks,based on the existing graph autoencoder,this paper proposes a novel adversarial residual graph variational autoencoder with batch normalization.Firstly,in order to make the KL divergence consistent with the dataset distribution,the approximate posterior parameters are batch normalized,which can ensure that the expected value of the KL distribution is positive,effectively avoiding posterior collapse.In addition,we invite adversarial network and residual connection into the model so as to enhance the expressive power of latent vectors and obtain a more stable embedding representation of the topology and content information of graph.The results of link prediction experiments on three reference datasets show that the average accuracy of the model proposed is higher than 93.3% and the AUC score is higher than92.5%,which is better than the current mainstream graph autoencoder.That proves the model can fully integrate topology information and node attribute information and realize the embedded representation of the diverse characteristics of the social network.3.Aiming at the problem of semantic shift in the process of online social network group discovery,this paper proposes a social network group discovery algorithm based on the local extension of seed groups.First,the social networks under different activity semantics are transformed into the embedded representation of low-dimensional vector space with variational graph autoencoder.Next,starting from the manually labeled seed groups,we iteratively expand on the multi-element network reconstructed based on three social network models with different activity semantics,dynamically update the group representation,and reduce the semantic offset.Finally,we discover the group related to public security in online social network.The experimental results on real social network data show that the accuracy,recall and F1 value of the group discovery algorithm proposed in this paper can reach the highest,which is at least 5% higher than the current mainstream group discovery algorithm;At the same time,the results also show that the multi-dimensional activity semantic extraction scheme used to construct multiple single-modal sub-networks todescribe social networks can significantly improve the discovery effect of the target group.Finally,this paper uses the above key technologies to develop a social network group discovery system based on activity semantics,which can automatically complete online social network data collection,data cleaning,activity semantic extraction,network model construction,social network graph embedding,and target group discovery.and other tasks to meet the needs of the field.