| Network alignment is a task that matches entities in different networks using network structure information and node attribute information.This task matches entities from different networks to integrate information from multiple networks,thus obtaining a more comprehensive and in-depth understanding.In social networks,network alignment can better understand user behaviors and interests,thus improving the quality of social network services and advertising effectiveness.Therefore,network alignment in social networks has important significance for research and applications in various fields.However,noise is inevitably introduced during the network construction process,which can affect the performance of network alignment.To mitigate the impact of noise on network alignment,many researchers have introduced representation learning techniques into network alignment tasks and achieved significant improvements.Although representation learningbased network alignment algorithms have made remarkable progress,there are still some limitations in handling attribute and structural information.This study aims to propose feasible improvement strategies for processing these two types of information and to explore their implementation methods and effects in depth.Specifically,the research work of this paper is as follows:1.Existing network alignment algorithms based on representation learning have limitations in handling attribute information,and cannot solve the problem of word polysemy in attribute texts.Therefore,this paper introduces pre-trained language models into the network alignment task and proposes the NABP(Network Alignment based Pre-trained Model)algorithm.The pre-trained language model is used to process attribute texts and mine the deep semantics of attribute texts.This chapter conducted experiments on two real social network datasets using the NABP algorithm and compared it with seven existing network alignment algorithms.The selection of attribute feature extraction models and parameters were also analyzed.The experiment shows that using pre-trained language models can better extract hidden information in texts and achieve higher accuracy compared to current algorithms.2.We propose a social network alignment algorithm,NABDAT(Network Alignment based Domain Adversarial Training),based on domain adversarial training to address two issues with the NABP algorithm.First,in combining network structure information and attribute information,NABP concatenates node embeddings that contain network structure information and attribute features,but this method fails to effectively fuse the two types of data.To address this issue,NABDAT uses graph convolutional networks as the representation learning method to integrate node attribute information and network structure information during representation learning.Second,NABP does not eliminate domain-specific information in the network.To tackle this issue,NABDAT optimizes the training process by using domain adversarial training.NABDAT uses graph convolutional networks as feature extractors and multilayer perceptrons as domain classifiers to filter out domain-specific information.Experimental results show that using graph convolutional networks combined with domain adversarial training as a representation learning method can effectively improve the performance of network alignment algorithms,especially on datasets with sufficient structural information and insufficient attribute information.3.The first two network alignment algorithms both require supervised label information,which requires a lot of manual effort to obtain.To eliminate the dependence on label information,this chapter proposes a new unsupervised network alignment algorithm called UNAA(Unsupervised Network Alignment Algorithm).Compared to existing unsupervised network alignment algorithms,UNAA solves the problem of node neighborhood consistency and maps the node representations of two networks to the same vector space using a linear mapping function.First,existing graph representation learning algorithms are used to learn representations of the source and target networks,obtaining node representations of the two networks.Secondly,to ensure that the node representations of the two networks are comparable while maintaining node neighborhood consistency,singular value decomposition and Sinkhorn algorithm are used to map the representation matrices of the two networks to the same vector space.In order to more accurately evaluate the performance of the algorithm,this paper uses simulation scenarios for experimentation.The experiments show that on sparse networks,the performance of the UNAA algorithm is better than other unsupervised algorithms,and on dense networks,the performance of the UNAA algorithm is also better than most baseline algorithms. |
