Abstract: In the application of social item recommendation, in addition to considering the social relationship between users, mining the implicit relationship features between items also plays a crucial role in enhancing the representation learning ability of users and items. However, existing social item recommendation methods constructed their item graphs using a static method, which was difficult to accurately capture the latent relationship between items, and the subsequent graph fusion process lacked deep interaction, which may limit the model's ability to understand the complex and multi-level relationship between the features of multiple graphs. To this end, a multi-task social item recommendation based on dynamic adaptive generation of item graph (IMTSR) is proposed. Firstly, in the joint training process based on multi-task learning, the item graph dynamic generation module adaptively adjusted the item graph structure by combining the feedback information of downstream recommendation tasks. Secondly, the social item recommendation module used a deep multi-graph feature fusion mechanism to iteratively propagate the feature representation of users and items among input graphs, so as to capture the complex correlation between multiple graphs more deeply. Finally, comparison experiments with six strong baselines on two public datasets show that IMTSR improved by at least 3% on each evaluation metric compared to the sub-optimal recommendation results. The robustness of IMTSR has been fully verified in the evaluation experiment for cold user and cold item recommendation, which indicates that IMTSR can alleviate the recommendation problem of cold user and cold item with sparse interactions to some extent.

Key words: social network, graph neural network, multi-task learning, graph representation learning, social item recommendation

摘要： 在社交项推荐应用中，除了考虑用户间的社交关系，挖掘项间隐含的关系特征对提升用户和项的表示学习能力同样具有至关重要的作用。然而，现有社交项推荐方法中静态的项图构建过程难以准确抓取项间的潜在关系，且后续的图融合过程缺乏深度交互，这可能限制了模型对多图特征间复杂、多层次关系的理解能力。为此，提出了一种基于项图动态适应性生成的多任务社交项推荐方法(IMTSR)。首先，在基于多任务学习的联合训练过程中，项图动态生成模块会结合下游推荐任务的反馈信息来适应性地调整项图结构。其次，社交项推荐模块采用深层次的多图特征融合机制在各输入图之间迭代地传播用户和项的特征表示，进而更深入地捕捉多图之间的复杂关联。最后，在两个公共数据集上与6种先进基线方法的对比实验表明，IMTSR相较于次优推荐结果在各项评估指标上均至少提高了3%。IMTSR的鲁棒性在针对冷用户和冷项推荐的评估实验中得到了充分验证，这表明IMTSR可以在一定程度上缓解交互稀疏的冷用户和冷项的推荐问题。

关键词: 社交网络, 图神经网络, 多任务学习, 图表示学习, 社交项推荐