Research On Robust Embedding For Attributed Graphs Via Topological Denoising

Attributed graph is a data structure with powerful representation capability,which can represent both the properties of objects themselves and the relationships between ob-jects,and is therefore widely used in scenarios such as social networks,recommendation systems,and financial engineering.Previous attributed graph mining algorithms often rely heavily on topology,and when the topology is affected by noise and becomes unre-liable,the quality of the embeddings is also severely affected.Therefore,the robustness of attributed graph embeddings has received increasing attention.However,current algo-rithms for attributed graph robust embedding have two main shortcomings:on the one hand,due to the invisibility of topology noise,it is impossible to classify the edges be-tween nodes by supervised learning,so most algorithms are based on heuristic methods to eliminate the negative effects of topology noise,which is difficult to capture topology noise efficiently.Meanwhile,if the intensity of denoising is increased,it is easy to damage the beneficial information in the topology.On the other hand,the training of most algo-rithms relies on joint learning with downstream tasks,and this approach is easily limited by downstream tasks.Currently,some algorithms try to remove the noisy information in topology structure by unsupervised learning,but these algorithms are often difficult to jointly use the attributed information and topology information of nodes,so it is difficult to obtain a high-quality topology structure.The main contributions of this thesis are as follows.First,to capture invisible topological noise,this thesis proposes the attack-assisted attributed graph denoising algorithm AGD from the perspective of attack.The AGD trans-forms attributed graph topology denoising into a supervised learning task by generating auxiliary attack samples as references,so as to effectively improve the quality of attributed graph topology.Meanwhile,based on the guidance of auxiliary attacks,the AGD is still able to remove the most harmful noise information in priority when the denoising ratio is low.Experiments based on 6 real-world graphs show that the node classification accuracy of AGD is improved by 11%compared with the original algorithm in the scenario with25%Mettack.Second,to avoid damage the beneficial information in topology when denoising,this thesis proposes the AGD~+algorithm,which carries out low-intensity denoising for topology structure,and then uses knowledge distillation technology to enable the model to output high-quality embedding only based on node features,thereby avoiding the in-terference of residual noisy information in the topology structure.Experiments based on6 real-world graphs show that the node classification accuracy of AGD~+is improved by16%compared with the original algorithm in the scenario with 25%Mettack.Third,to overcome the difficulty of integrating attribute and topology information in unsupervised denoising,this thesis proposes a hidden embedding-driven denoising algo-rithm LFGD for attributed graphs.LFGD can give the hidden embedding information of nodes to efficiently approximate the low-rank matrix decomposition,powerfully remove the topology noise belonging to higher-rank information,and improve the quality of at-tributed graph topology without relying on the downstream task.Experiments show that the LFGD is effective against many types of attacks and has a time performance advan-tage.In the Pub Med dataset,the node classification accuracy of LFGD fluctuates less than3%when facing three types of attack methods.