Research On Critical Algorithms For Cognizing Social Network Entities

Along with the maturity of Internet technology,social networks have shown unique value on social,political and economic aspects,thus cognitive algorithms for social net-work entities have become one of the research focuses in computer science.The previous studies play a pivotal role in online services,online marketing and public security.How-ever,there is still a substantial amount of unsolved technical problems.This dissertation concentrates on such problems and investigates cognizing the entities on social networks,thus containing the high academic and application value.Specifically,we focus on four cruxes of cognizing entities,including:(1)prevalent time prediction of topics;(2)user opinion prediction;(3)key user search based on influence relationship;(4)opinion per-suasion based on key user search.The detailed introduction of our works are as follows.(1)Text-enhanced prevalent time prediction of topics.The previous works mainly suffer from the low prediction accuracy,since they fail to utilize the multi-view infor-mation of topics and learn complex varying patterns of topic prevalent time.In order to improve the prediction accuracy,a text-enhanced deep survival model is proposed,which includes a text-enhanced module in the front and a deep survival module in the back.To fully utilize multi-view topic-related information,the front text enhancement module con-denses the text information into topic-related key features,thus utilizing topic-related text information.To effectively learn complex varying patterns of topic prevalent time,the posterior deep survival module adopts the theoretical framework of survival model.It fully explores the temporal relevance of input information through its deep recurrent neu-ral network structure,and uses a special loss function to effectively learn complex varying patterns of topic prevalent time.Extensive evaluations on different real-world datasets re-veal that the proposed model outperforms the state-of-the-art prevalent time prediction model and achieves accurate prediction of topic prevalent time.(2)User opinion prediction with multi-view information fusion.Existing works mainly suffer from the low prediction accuracy,which is mainly because of the difficulties of exploiting multi-view user information and learning diverse user preferences.To im-prove the prediction accuracy,we propose a disentangled user representation model with multi-view information fusion.This model includes a multi-information adversarial fu-sion method and a disentangled variational auto-encoder.To fully utilize multi-view user information,we propose an adversarial fusion method to fuse grouped topic correlation,user historical opinions and user relationships.To effectively learn the user preference,we design a disentangled variational auto-encoder to disentangle latent factors that dominate user preferences but may entangle with each other.Extensive evaluations on different real-world datasets reveal that the proposed model performs better than the comparisons and achieves accurate prediction.(3)Second-order key user search.The main reason for the low accuracy of existing key user search methods is the high complexity of modeling user influence relationships,thus rendering it difficult to accurately estimate user influence.In order to improve the accuracy of key user search,we first define the second-order influence,and then propose a second-order independent cascade model that integrates the first-order and second-order influence.To address the challenge of user influence estimation,the model adopts an innovative design of second-order influence propagation rules to overcome the biased in-fluence estimation problem of existing models.As considering the second-order influence increases the complexity of influence estimation,a second-order key user search algorithm and its parallel acceleration algorithm are proposed based on the inverse influence sam-pling method.The algorithms are both accurate in discovering key users and applicable to large-scale social networks.Extensive evaluations on different real-world datasets reveal that the proposed model is better than the most widely used first-order influence propaga-tion model.The proposed algorithm is more effective and efficient to achieve the accurate key user search.(4)Opinion persuasion based on key user search.The existing research mainly suf-fers from the problem of low accuracy of key user search,which is mainly because of the high complexity of the target problem,making it difficult to design effective algorithms.In order to improve the accuracy,we design an optimal strategy for determining user psychological features,and propose an improved greedy algorithm based on the previous research.Secondly,with respect to large-scale social networks,we propose a community-based algorithm for acceleration,which improves the efficiency with the premise of sacri-ficing very little effectiveness.Based on the approximate submodularity,the approximate ratios of the proposed algorithms are theoretically demonstrated.Extensive evaluations on different real-world datasets reveal that the proposed algorithms outperform the state-of-the-art algorithms.The community-based algorithm is both accurate and applicable to large-scale social networks.In this dissertation,the proposed methods are validated based on public experimental datasets.The evaluation results fully demonstrate that the proposed methods effectively solve the corresponding key challenges,and the proposed methods are effective techniques for solving the cognizing tasks of social network entities.