Research On Network Representation Learning Based On Side Information Extraction

The explosive growth of data in information networks has brought huge challenges to data analysis and network information exploration.Network representation is a promising way to solve these problems.Network representation learning methods,which are considered as the basis for network exploration,encode network into a low-dimension and dense vector space by preserving valuable information automatically.However,most of the methods only consider topological information and cannot explore edge or attribute information in the network deeply,resulting in poor performance at downstream tasks such as classification,clustering and link prediction.In order to solve these problems,a novel way to extract network information is proposed.Based on this,a representation learning model for attribute networks is further proposed.The detailed work contents are shown below.Most of the existing methods ignore edge representation vectors in the network and use edge information insufficiently.In order to solve these problems,a network representation learning method based on edge information extraction is proposed.First,the original network is transformed into an edge network with structure and edge information.Then,edge representation vectors can be obtained directly by using an existing network representation model with edge network as its input.Extensive experiments on several real-world networks demonstrate that edge network outperforms original network in various graph mining tasks.In order to use edge information of attribute network better,AE-MVANR,a multi-view attribute network representation learning method based on auto-encoder are proposed.First,structure information is transformed into the structure-view network and the attribute-view network is constructed by calculating the co-occurrence frequency of the same attributes between nodes.Then,a random walk algorithm is used to obtain a series of node sequences on two views.At last,by inputting node sequences into an auto-encoder,node representation vectors that ingrate structure and attribute information can be obtained.Extensive experiments on several real-world networks demonstrate that AE-MVANR outperforms the state-of-the-art techniques based on only structure information or only attribute information in various network mining tasks,i.e.,node classification and node clustering.