融合多模态自监督图学习的视频推荐模型

现有视频推荐方法在算法框架中引入图神经网络来建模用户—视频协同关系，学习用户和视频的表示向量，但是节点中包含的冗余噪声会限制模型的建模能力。针对以上问题，提出了一种融合多模态自监督图学习的视频推荐模型(IMSGL-VRM)。首先，在图数据增强模式下构建自监督的图神经网络模型学习多模态视图下的节点特征表示，以提升节点表示的泛化能力；其次，为了得到推荐结果的多样性，设计了多兴趣提取模块从用户历史的交互视频序列中建模用户的多兴趣；最后，融合多模态的用户多兴趣表示和视频的特征表示，使用多样性可控的方式输出推荐结果，以满足视频推荐的多样性需求。在MovieLens-1M和TikTok数据集上实验，采用准确性、召回率、NDCG和多样性等指标评估模型。实验结果表明，该模型相比经典基准模型均有显著的性能提升。

Self-supervised graph learning of fusing multi-modal for video recommendation model

Existing video recommendation methods introduce graph neural networks in the framework to model the user-video co-relation and learn the representation vectors of users and videos, but the redundant noise contained in the nodes may limit the modeling capability of the model. To this end, this paper proposed a new model that integrated multimodal self-supervised graph learning for video recommendation model(IMSGL-VRM). First, this paper constructed a self-supervised graph neural network in the graph data augmentation mode to learn the node feature representation in the multimodal view to improve the generalization ability of the node representation. Second, in order to obtain the diversity of recommendation results, this paper designed a multi-interests extraction module which models users'' multi-interests from their historical interactive video sequences. Finally, this paper integrated the multi-modal users'' multi-interest representation and the representation of video''s feature and obtained recommendation results in a controllable way to satisfy the diversity requirement of video recommendation. This paper conducted experiments on MovieLens-1M and TikTok datasets, and evaluated the model performance using accuracy, recall, NDCG and diversity metrics. The experimental results show that IMSGL-VRM has significant performance improvement compared with the classical benchmark model.