关键词: 自适应门控机制, 下一个POI推荐, 图神经网络

Abstract: The recommendation of the next POI has attracted much attention in location-based social networks,aiming to accurately recommend POIs based on users' historical check-ins and temporal information.However,traditional methods fail to consider temporal sequences and have low learning efficiency for graph nodes.To address this,in this paper,we propose a method for next POI recommendation by incorporating adaptive gating mechanisms into geographic and sequential graph modules.The proposed network consists of three main parts: the adaptive geographic graph module,which combines the adaptive gating mechanism with graph convolutional neural networks to adjust node fusion update weights by using gating signals; the adaptive sequential graph module,which learns user access preferences through a random walk network and enhances the weights of relevant preferences on the basis of target task attributes using adaptive gating mechanisms; and the semantic joint module,which maximizes the consistency distribution between the geographic and sequential graph modules and optimizes the loss of the joint framework via soft-label cross-entropy loss functions.To validate the effectiveness of the model,experiments are conducted on foreign datasets (Foursquare_NYC and Foursquare_TKY)and a domestic dataset(Microblog).The experimental results demonstrate that the proposed model achieves recommendation accuracies of over 85% across all datasets,with performance improvements ranging from 2.97% to 86.90% over those of state-of-the-art baseline models.

Key words: adaptive gating mechanism, next POI recommendation, graph neural networks