Research On Network Representation Learning Algorithms Based On Multi-Information Fusion

With the development of information technology and the advent of the era of big data,the scale of data in real life has achieved a dramatic growth,and data are further correlated with each other based on related association rules,which form various types of networks.How to mine potential complex association patterns from network data and build evolutionary prediction model is of great significance to the development of society.With the continuous growth of the network scale,the traditional network analysis methods cannot be flexibly used due to the limitation of time and space computation costs.Network representation learning aims to embed nodes into the low-dimensional vector representation space,and has attracted extensive attention from researchers because of its outstanding network representation ability in the field of network analysis.The real-world networks usually contain multiple information;however,previous network representation learning algorithms mainly focus on single information modeling or suffer from insufficient multiple information mining,which degrades the performance of the network representation learning algorithms.In view of this,this dissertation will conduct research on network representation learning algorithm based on multi-information fusion,aiming to explore more effective multi-information fusion and model construction methods.Comprehensive mining and utilization of multi-information in the network can generate more expressive network representation and significantly improve the quality of node representations.Specifically,the main research contents and contributions of this dissertation are summarized as follows.(1)We propose a novel network representation learning algorithm,which effectively preserves the high-order node proximity and incorporates the hidden community structure into node representations.Based on the assumption that nodes within the same community and nodes with higher proximities should have similar vector representations,we propose to adopt non-negative matrix decomposition technique to embed the high-order node proximity and community structure into a single low-dimensional representation space,and then introduce the low-dimensional community vector representation to bridge the connection between them.As a consequence,a multi-structure information fusion representation model is established,and an efficient alternating optimization strategy is further proposed to solve the non-convex optimization problem in this model.Our proposed algorithm can learn more expressive node representations,whose effectiveness has been effectively verified in network analysis tasks such as node classification,network reconstruction and link prediction.(2)We propose a novel network representation learning algorithm for dual semantic information fusion under node semantic conflict.To be specific,we propose to use two vector representations to represent the different roles and conflicting semantics of nodes,and introduce a novel attention mechanism to estimate the coefficients between nodes under different semantics.Based on the balance theory,we further propose a novel graph attentional neural network framework to effectively capture the complex semantic interaction between nodes and fuse the conflicting semantics of nodes.Besides,a novel objective function is also designed to optimize this framework.The effectiveness of our proposed algorithm is verified in link sign prediction task of real-world signed networks.(3)We propose a novel network representation learning algorithm based on the fusion of low-frequency information and high-frequency information to effectively embed both the similarity and dissimilarity between nodes into a single vector representation space.Based on the assumption that similar and dissimilar connected nodes should be close or distant respectively in the representation space,we propose to use the low-frequency information and high-frequency information in networks to preserve the similarity and dissimilarity between connected nodes with the help of spectral graph theory and graph signal processing.To be specific,we design customized low-frequency graph convolution filters and high-frequency graph convolution filters to extract low-frequency information and high-frequency information,and then combine all the graph convolution filters into a unified message passing framework to effectively fuse low-frequency information and high-frequency information.In addition,we propose a novel self-gating mechanism to estimate the impacts of low-frequency information and high-frequency information during message passing.The experimental results on link sign prediction task of real-world signed networks demonstrate the effectiveness of our proposed algorithm,which significantly outperforms the state-of-the-art methods and achieves significant performance improvement.