We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
基于图神经网络的兴趣点推荐的隐私保护框架.
- Authors
江欣俞; 李晓会; 秦若婷; 张爱
- Abstract
The traditional research on the interest point model based on graph neural network is to define the weight through a simple attention mechanism, or simply linearly combine various factors, and lack the semantic information and interaction information that considers the user and the interest point itself from multiple perspectives. In addition, the existing graph neural network recommendation relies on the centralized storage and training of graph structure information, which has the risk of privacy leakage. In order to solve the above problems, a privacy-preserving framework for point of interest(POI) recommendations for graph neural networks (PPGNN) based on graph neural networks was proposed. Firstly, the graph structure was strengthened by introducing multi-feature pattern and attention mechanism, and the graph model of user social relationship was constructed. Secondly, the sampling algorithm of neighbor nodes of interest points and the convolutional aggregation mechanism were redesigned from the perspective of multiple scenes, and the semantic-level attention mechanism was used to aggregate heterogeneous graphs. Finally, a variable dynamic gradient client differential privacy algorithm was proposed to achieve the effect of both optimization and feedback. A large number of experiments were conducted on different data sets of Yelp and Gowalla to prove the effectiveness of this scheme, which makes up for the limitations of graph neural network recommendation due to privacy threats, and is superior to centralized graph neural network recommendation method and traditional point of interest recommendation method. In addition, PPGNN can better overcome the problems of data sparse and cold start in recommendation.
- Publication
Science Technology & Engineering, 2023, Vol 23, Issue 17, p7407
- ISSN
1671-1815
- Publication type
Article