近年来, 社交推荐的研究主要聚焦于社交网络中显式、隐式关系的联合建模, 却忽视了高阶隐式关系并非对每个用户都同等重要这一特殊现象. 高阶隐式关系对一个有着足够多邻居的用户与一个仅有少量邻居的用户重要性存在明显差异. 此外, 由于社交关系建立的随机性, 显式关系并不总是可用的. 提出了一种新的自适应高阶隐式关系建模方法(adaptive high-order implicit relations modeling, AHIRM), 该模型由3个部分组成: 首先, 过滤不可靠关系且识别出潜在可靠关系. 旨在避免不可靠关系带来的负面影响, 并部分缓解数据稀疏的问题; 其次, 设计自适应随机游走算法, 结合规范化后的节点中心度为用户捕获不同阶数的邻居, 构建用户间的高阶隐式关系, 进而重构社交网络; 最后, 运用图卷积网络(graph convolutional network, GCN)聚合邻居节点信息, 更新用户嵌入, 实现高阶隐式关系建模, 从而进一步缓解数据稀疏问题. 在建模过程中, 同时考虑到社交结构和个人偏好的影响, 模拟并保留了社交影响传播的过程. 在LastFM、Douban和Gowalla这3个数据集上与相关算法做了对比验证, 结果证实了该模型的有效性和合理性.
Recent research studies on social recommendation have focused on the joint modeling of the explicit and implicit relations in social networks and overlooked the special phenomenon that high-order implicit relations are not equally important to each user. The importance of high-order implicit relations to users with plenty of neighbors differs greatly from that to users with few neighbors. In addition, due to the randomness of social relation construction, explicit relations are not always available. This study proposes a novel adaptive high-order implicit relations modeling (AHIRM) method, and the model consists of three components. Specifically, unreliable relations are filtered, and potential reliable relations are identified, thereby mitigating the adverse effects of unreliable relations and alleviating the data sparsity issue. Then, an adaptive random walk algorithm is designed to capture neighbors at different orders for users according to normalized node centrality, construct high-order implicit relations among the users, and ultimately reconstruct the social network. Finally, the graph convolutional network (GCN) is employed to aggregate information about neighbor nodes. User embeddings are thereby updated to model the high-order implicit relations and further alleviate the data sparsity issue. The influence of social structure and personal preference are both considered during modeling, and the process of social influence propagation is simulated and retained. Comparative verification of the proposed model and the existing algorithms are conducted on the LastFM, Douban, and Gowalla datasets, and the results verify the effectiveness and rationality of the proposed AHIRM model.