Abstract:

To address the data sparsity problem in multimodal recommendation algorithms and the problem in the existing Self-Supervised Learning （SSL） algorithms that the algorithms often focus on SSL a single feature in a dataset， ignoring the possibility of joint learning of multiple features， a multimodal fusion recommendation algorithm based on joint self-supervised learning was proposed， called SFELMMR （SelF-supErvised Learning for MultiModal Recommendation）. Firstly， the existing SSL strategies were integrated and optimized to enhance data representation capabilities significantly by learning data features from different modalities jointly， thereby alleviating the data sparsity issue. Secondly， a method to construct multimodal latent semantic graph was designed by integrating deep item relationships from a global perspective with direct interactions from a local perspective， enabling the algorithm to capture complex relationships among items more accurately. Finally， experiments were carried out on three datasets. The results demonstrate that the proposed algorithm achieves significant improvements in multiple recommendation performance metrics compared to the existing best-performing multimodal recommendation algorithms. Specifically， the proposed algorithm has the Recall@10 improved by 5.49%， 2.56%， and 2.99%， respectively， the NDCG@10 improved by 1.17%， 1.98%， and 3.52%， respectively， the Precision@10 improved by 4.69%， 2.74%， and 1.22%， respectively， and the Map@10 improved by 0.81%， 1.59%， and 3.11%， respectively. Besides， through ablation experiments of the proposed algorithm， the effectiveness of the algorithm is verified.

Key words: recommendation system, multimodal, Self-Supervised Learning (SSL), Graph Convolutional neural Network (GCN), feature fusion

摘要：

针对多模态推荐算法的数据稀疏性问题，以及现有的自监督学习（SSL）算法往往集中在对数据集中单一特征的SSL上，而忽视了多特征联合学习的可能性的问题，提出一种基于联合SSL的多模态融合推荐算法SFELMMR （SelF supErvised Learning for MultiModal Recommendation）。首先，整合并优化现有的SSL策略，以通过联合学习不同模态的数据特征，显著增强数据的表示能力，从而缓解数据稀疏性的问题；其次，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，设计一种构造多模态潜在语义图的方法，使算法能更精准地捕捉项目间的复杂联系；最后，在3个数据集上进行实验。结果表明，与现有算法中表现最佳的多模态推荐算法相比，所提算法在多个推荐性能指标上取得了显著提升。具体地，所提算法的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%。此外，通过对所提算法进行消融实验，验证了该算法的有效性。

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