【摘要】 为解决基于图神经网络社交推荐方法只考虑用户的社交关系，忽略了项目之间的相互依赖关系，以及没有考虑用户和项目的偏好偏移的问题，提出一种基于分布式和图注意力机制的社交推荐算法。利用图神经网络获得用户和项目之间的关联，得到用户和项目之间的依赖关系；把用户和项目的偏差看作向量，融合到学习用户和项目表示的过程中，利用分散邻域集捕获统计偏差偏移；结合多头注意力机制关注全局信息的权重分配，提升特征提取能力。在对Ciao和Epinions这两个真实数据集进行测试的过程中，所得结果显示，GNN-DMR模型在均方根误差和平均绝对误差这两个评价指标上，相较其它算法均展现出更低的数值，验证了该推荐系统模型的高效性和实用性。更多还原

【Abstract】 To solve the problem that the graph neural network-based social recommendation method only consider the user’s social relationship and ignores the interdependence between items, and does not consider the preference shift of users and items, a social recommendation algorithm based on distribution and graph attention mechanism was proposed. The graph neural network was used to obtain the correlation between users and projects, the dependencies between users and projects were obtained. The bias of users and items was treated as vectors, they were integrated into the process of learning the representation of users and items, and the scattered neighborhood set was used to capture the statistical bias shift. The multi-head attention mechanism was combined to pay attention to the weight allocation of global information to improve the feature extraction ability. In the process of testing the two real datasets of Ciao and Epinions, the results show that the GNN-DMR model shows lower values than other algorithms in the two evaluation indexes of root mean square error and mean absolute error, which verifies the efficiency and practicability of the recommended system model.更多还原