User decision-making behavior in recommender systems is jointly driven by a large number of underlying factors. Learning and revealing the representations of these latent factors can provide more robustness and interpretability. However, mining the latent intentions of user decision-making behaviors in existing multimodal recommendation studies faces the following two key challenges: (i) Modal Noise Pollution : In multimodal user intent modeling, inputs from individual modalities are inevitably corrupted by noise of varying severity. During message passing, a large proportion of irrelevant or even contradictory signals are propagated and injected into item representations, which impedes the model’s ability to achieve pure semantic alignment at the content level. (ii) User Intent Confounding : Real-world items naturally possess multiple attributes, with different attributes of the same item influencing distinct potential user intentions. However, in existing modeling designs for user intent, such intents are mapped onto user–item interaction labels of the same coarse granularity. This many-to-one mapping relationship between intentions and items leads to significant confounding and dilution of users’ fine-grained intentions. To address the above challenges, this work pays special attention to the implied user intent behind pure multimodal features. Specifically, we construct a dynamic, adaptive, multimodal intent disentanglement model. This model adopts a non-ID paradigm and mines the distribution of user intents in multimodal scenarios directly from users’ decision-making behaviors. A comprehensive experimental study on the Amazon dataset shows that the method is effective and provides a novel learning scheme for mining user intent in multimodal scenarios.

Self-supervised tasks show significant advantages for node representation learning in recommender systems. This core idea of self-supervised task-based recommender systems depends on data augmentation to generate multi-view representations. However, there are two key challenges that are not well explored in existing self-supervised tasks: i) Restricted by the structure of the graph-based CF paradigm itself, the classical graph comparison learning architecture ignores the global structural information on the user-item interaction graph. ii) In a key part of existing contrast learning-random graph data enhancement schemes can significantly deteriorate model performance. To address these challenges, we propose a new hypergraph collaborative filtering with adaptive augmentation framework(HCFAA). It captures both local and global collaborative relationships on the user-item graph through a hypergraph-enhanced joint learning architecture. In particular, the designed adaptive structure-guided model ignores the noise introduced on unimportant edges, and thus learns the critical node information on the user-item graph. Comprehensive experimental studies on the Amazon dataset show that the method is effective, which provides an optimization scheme with a new perspective for the problems of key node loss in graph data enhancement and loss of higher-order structural information in GNN. The source code of our model can be available onhttps://github.com/RSnewbie/RS/tree/master/HCFAA.