Abstract: Abstract: To address the data sparsity problem in multimodal recommendation algorithms and the limitation of existing self-supervised learning algorithms that often focus on learning a single feature in a dataset, ignoring the potential of joint learning of multiple features, this paper proposes a multimodal fusion recommendation algorithm based on joint self-supervised learning. First, we integrate and optimize existing self-supervised learning strategies to significantly enhance data representation capabilities by jointly learning data features from different modalities, thereby alleviating the data sparsity issue. Second, we design a method to construct a multimodal latent semantic graph by integrating deep project relationships from a global perspective with direct interactions from a local perspective, enabling the algorithm to more accurately capture complex relationships among projects. Experimental results on three datasets demonstrate that the proposed algorithm achieves significant improvements in multiple recommendation performance metrics compared to the state-of-the-art multimodal recommendation algorithms. Specifically, Recall@10 improves by 5.49%, 2.56%, and 2.99%, NDCG@10 improves by 1.17%, 1.98%, and 3.52%, Precision@10 improves by 4.69%, 2.74%, and 1.22%, and Map@10 improves by 0.81%, 1.59%, and 3.11%. Furthermore, ablation experiments on the proposed algorithm verify the effectiveness of the method.

Key words: Keywords: recommendation system, multimodal, self-supervised learning, graph convolutional network, feature fusion

摘要： 摘 要: 针对多模态推荐算法的数据稀疏性问题，以及现有自监督学习算法往往集中在对数据集中单一特征的自监督学习上，忽视了多特征联合学习的可能性，提出了一种基于联合自监督学习的多模态融合推荐算法。首先，整合并优化现有的自监督学习策略，通过联合学习不同模态的数据特征，显著提升了数据的表示能力，从而缓解了数据稀疏性问题。其次，设计了一种构造多模态潜在语义图的方法，通过融合全局视角下的深层次项目关系和局部视角下的直接相互作用，使得算法能够更加精准地捕捉项目间的复杂联系。通过在三个数据集上的实验结果表明，与现有算法中表现最佳的多模态推荐算法相比，提出的算法在多个推荐性能指标上取得了显著提升，Recall@10分别提升了5.49%、2.56%、2.99%，NDCG@10分别提升了1.17%、1.98%、3.52%，Precision@10分别提升了4.69%、2.74%，1.22%，Map@10分别提升了0.81%、1.59%、3.11%。并且通过对本文提出的算法进行消融实验，验证了方法的有效性。

关键词: 关键词: 推荐系统, 多模态, 自监督学习, 图卷积神经网络, 特征融合