Research On Influence Maximization Based On Multilayer Network Representation Learning

With the continuous development of computer technology,related studies on complex networks have attracted the attention of a large number of scholars.Among them,influence maximization problem has important theoretical value and practical significance for public opinion control and epidemic prevention and control,so it has become one of the hot spots in the field of complex networks.The influence maximization problem refers to mining a certain number of seed nodes in a specific propagation model so that its influence can be effectively and widely spread.However,the existing research results on influence maximization problem mainly focus on a single homogeneous network and capture the network topology by introducing heuristic pruning to find the node sets with maximum influence.In the real world,node users often belong to multiple types of networks at the same time,and all types of networks influence each other and depend on each other.There are two forms of modeling for many types of networks,including single heterogeneous network and multilayer network.Since the single heterogeneous network is still a single-layer network,and the structure information between the layers is ignored.Therefore,it is of practical significance to study the influence maximization of nodes in a multilayer network model with multiple information.In addition,limited to pruning strategy,the classical method will inevitably lose part of seed information in the selection process of the seed set,and cannot be directly applied to multilayer networks,resulting in reduced effectiveness of the method.To solve the above problem,this thesis introduces the network representation learning technology,which can learn the feature representation of the whole multilayer network,and makes up for the disadvantage of the traditional influence maximization method losing part of the information in the process of the seed set selection.So far,this thesis proposes an influence maximization method based on multilayer network representation learning,and the specific research work is as follows:(1)In multilayer network representation learning,in order to obtain the corpus containing network structure features,this thesis uses random walk to traverse nodes between and within layers of a multilayer network.Firstly,the traditional network representation learning model lacks the learning of local structural features of nodes.Therefore,in order to enable the network representation learning model to learn rich network information,this thesis uses the number of common neighbors to describe the similarity among users,and uses the clustering coefficient to describe the transmission of information among users.The Cluster Rank method is used to fuse the above two indicators,and then a node structure indicator is proposed to guide the random walk.Secondly,a cross-layer parameter is introduced to carry out random walk between network layers.Finally,the node sequence is formed according to the above random walk strategy,and the Skip-Gram model is used to maximize the likelihood probability of the random walk sequence.Based on the basic concept of multilayer network and the above two random walk strategies,a multilayer network representation learning method IFMNE based on random walk of multiple information is proposed.Experimental results are performed on five real multilayer networks,and the embedding vectors were evaluated by link prediction task.The accuracy was significantly improved on the basis of low time complexity,which further proves the feasibility of guiding random walk.(2)On the basis of the above research,the IFMNE model is used to represent nodes in the network,and the embedding vector of nodes is clustered.After the cluster center is obtained,the corresponding nodes are used as candidate seed set.Then,the greedy strategy is used to select the most influential nodes from the candidate seed set.So far,an influence maximization method MNEIM based on multilayer network representation learning is proposed,which can be divided into two methods according to the candidate seed node screening strategy,namely MNEIM-1 method and MNEIM-2 method.Firstly,MNEIM-1uses the K-means algorithm to cluster the node embedded vectors to obtain K clustering centers.Since the clustering centers obtained by the K-means algorithm may not have corresponding nodes in real datasets,the K nearest nodes are searched by the KD tree as candidate seed set.MNEIM-2 directly uses the DPC algorithm to select the top K central nodes as candidate seed set.Then,the traditional greedy algorithm CELF is used to screen the seed set.Finally,the validation on six real datasets of multilayer network shows that compared with the traditional influence maximization method,the MNEIM-1 method is superior to the existing heuristic algorithm in the propagation range,and has higher time efficiency than the greedy algorithm,which further proves the feasibility of the idea of introducing a multilayer network representation learning method.In the MNEIM method,the MNEIM-1 method has better propagation effect and expansibility than the MNEIM-2 method,but the MNEIM-2 method still has some advantages in small-scale datasets.