Study On Meta-learning Based Link Prediction For Multilayer Networks In Few-shot Settings

Link prediction in multilayer networks aims to predict the existence likelihoods of unobserved links at the target layer by fusing structural information at the auxiliary and target layers.Most existing methods assume the multilayer network already contains sufficient known structural information at each layer.However,real-world multilayer networks commonly have numerous nodes and complex relationships.Therefore,it is time-consuming and laborious to label sufficient structural information to support the prediction performance of existing methods.In commonly cases where the structural information at each layer of the multilayer network is insufficient,existing mainstream methods fail to achieve satisfactory prediction results.Thus,this thesis proposes two models based on meta-learning for link prediction in multilayer networks in few-shot settings.The first employs meta-learning to extract task-shareable knowledge adapted to the link prediction tasks at different layers in a multilayer network,and uses the extracted task-shareable knowledge as transferable knowledge to assist link prediction at the target layer;the second one significantly enhances the link prediction performance by introducing multiple attention mechanisms in the feature encoding process of links,which aims to mine the most relevant structural information for link formation.The main research of this thesis is as follows:(I)This thesis proposes the Task-shareable Knowledge Fusion Model(TKFM)to improve link prediction performance in multilayer networks in few-shot settings.TKFM comprises three primary components: the task divider,the shareable knowledge extractor,and the link predictor.The task divider divides link prediction tasks at auxiliary layers of the multilayer network into several few-shot link prediction tasks.The shareable knowledge extractor extracts task-shareable knowledge from the divided few-shot link prediction tasks on different network layers using the meta-learning mechanism.The link predictor uses the fine-tuning mechanism to fuse the task-shareable knowledge and the task-specific knowledge of the link prediction task at the target layer,and outputs the final prediction result.This thesis compares the performance of TKFM with eleven baselines.Experimental results on two real-world multilayer networks show that TKFM outperforms all baselines,demonstrating the effectiveness of the proposed model.Additionally,further analytical experiments confirm that TKFM is more stable than the baselines in few-shot settings.(II)To further improve link prediction performance in multilayer networks in fewshot settings,this thesis proposes the Combining Multiple Attention Mechanism and Taskshareable Knowledge Fusion Model(MATK)as an extension of the previously proposed TKFM.MATK introduces multiple attention mechanisms to differentiate the importance of structural information at different network layers when encoding feature representations for links.This enables MATK to mine the most relevant structural information for link formation.Afterwards,MATK uses meta-learning and fine-tuning mechanisms to efficiently fuse the task-shareable knowledge of link prediction tasks at different layers with the task-specific knowledge of the link prediction task at the target layer.This thesis selects twelve baselines,including TKFM,to compare performances.Experimental results on two real-world multilayer networks demonstrate the superiority of the proposed model MATK.Furthermore,a series of further analytical experiments confirm that combining multiple attention mechanisms and the meta-learning mechanism can effectively improve link prediction performance in multilayer networks in few-shot settings.