| The research of network data has become one of the hottest point questions of the data mining field in recent years.At present,the machine learning algorithm is based on inputting structured data,which is hard to be used in the network data directly.But,the appearance of network representation learning has provided another feasible way to solve this challenge.And network representation learning will excavate the deeper level of the sentence on the network of through abstract the raw network data layer-by-layer into the final feature representation required by the task,which will be available relieve the problem of network sparseness.The node learning process of network representation learning is coalesced with attribute information,which is beneficial to further excavate semantic information in the network to improve the final representation quality.However,this kind of algorithm ignored the priori information readily available in the data that make something already learned is a lack of discrimination.So,this article centered on the research between the priori information readily available in the data and the attribute network represents the fusion of learning,the main content as following:At first,the problem is lacking discrimination that in allusion to the nodes in ASNE algorithm represent the learning process,which already existed,did not take advantage of the priori information readily available in the data,then put forward the improved method,which is blending tag information in node learning process.This method is using a small number of supervision pieces of information to guide the learning process of node represent,which can recover the potential structure of the network Hence,the experiment was held in the two artificial networks and four real networks,and the result indicated that this method compared with the original algorithm,NMI and ACC improved by about 5%.On the other hand,aiming at the existing problems in the first experiment of attributive networks represents learning,such as the problems of the choice of node dimension reduction,attribute information and structural information can not perfectly mix and compare with label information constraints are more accessible to information.So,the improved method,which is adding preliminary training and optimized training has been provided.This kind of method enhanced the accuracy rate in the preliminary training through a newly constructed directed graph that focuses on optimizing the fusion of attribute information and structure information.It is not only using the pair constraint information that is easier to obtain to optimize the learning process,but also using the manner of data enhancement to reduce the computation and the time overhead.In simple terms,this method is mainly through optimizing the training process and enhancing the neighborhood structure information to improve the effect that brings by the random walk in preliminary training.And in the end,the experiment indicated that this kind of method was superior to the algorithm of merge tag information as we mentioned.Last but not least,to explore the practical value of the method as we mentioned,Python and Scrapy have been used to explore the user structure information from the Zhihu website and user profile to establish the real social network structure.Besides,in order to achieve the purpose of recommending potential friends to the following users through all kinds of methods,such as data pretreatment,to remove the noise data and the algorithm in this paper is used for link prediction. |
PDF Full Text Request