Multimodal recommendation systems have received widespread attention in recent years. The existing methods mainly treat multimodal features as a supplement to collaborative ID features to build alignment representation. However, these methods only use the user-item collaborative relation and item-item latent relation to implicitly represent multimodal features without explicitly using multimodal content, which ignores the intrinsic semantic relations within different modal content and makes the model not pay insufficient attention to user preferences, causing not great recommendation performance. Considering this problem, we argue the complete fine-grained representation of multimodal content, which includes collaboration relation, latent relation, and semantic relation. To address this, this paper proposes the Joint Content Semantic relation learning with Mamba for Multimodal Recommendation (JCSMRec). First, we construct Semantic relation representation with Mamba (SR2M) Module to explore the inter-modal correlation and provide an excellent aligned explicit representation for multimodal content. Then, we design a User preference-aware Module, according to user preference to guide learning the representation granularity of different modalities in the latent relations. Then, we conduct a joint representation to align and fusion multimodal features, which use the cyclic KL loss to align different semantic spaces and alleviate two joint losses to let user preference guide the learning process of the SR2M Module. This design effectively avoids over-reliance on ID features and enables a more comprehensive learning of textual and visual semantic information suitable for recommendation. Our method has achieved good results on multiple datasets and can be effectively inserted into other recommendation methods to improve results.

Recently, diffusion model-based methods have utilized user interest features as guidance conditions to achieve stable generation results in sequential recommendation tasks. However, these models struggle to capture users' dynamic interests, as the interests of different users are often inconsistent. Moreover, the fixed number of interests predefined by existing models cannot adapt to the diverse preferences of users, making it difficult to further improve recommendation performance. To address these issues, we propose a novel generative sequential recommendation framework named ADIGRec (Adaptive User Dynamic Interest Guidance for Generative Sequential Recommendation), which adaptively focuses on users' dynamic interest features. Specifically, our framework combines users' dynamic features and inherent interest features encoded from historical sequences as new guidance conditions. Furthermore, we introduce a module that injects dynamic interest features into the noise item embeddings, enabling explicit interaction with the guidance conditions during the generation phase. This approach essentially fits the noise in the target space rather than the user preference space, leading to improved recommendation diversity. Additionally, we propose a novel regularization method to mitigate the impact of user interest routing collapse on the generation results. Extensive experiments on three publicly available datasets demonstrate that our method achieves superior performance compared to established baseline methods.

Recent methods utilize graph contrastive Learning within graph-structured user-item interaction data for collaborative filtering and have demonstrated their efficacy in recommendation tasks. However, they ignore that the difference relation density of nodes between the user- and item-side causes the adaptability of graphs on bilateral nodes to be different after multi-hop graph interaction calculation, which limits existing models to achieve ideal results. To solve this issue, we propose a novel framework for recommendation tasks called Bilateral Unsymmetrical Graph Contrastive Learning (BusGCL) that consider the bilateral unsymmetry on user-item node relation density for sliced user and item graph reasoning better with bilateral slicing contrastive training. Especially, taking into account the aggregation ability of hypergraph-based graph convolutional network (GCN) in digging implicit similarities is more suitable for user nodes, embeddings generated from three different modules: hypergraph-based GCN, GCN and perturbed GCN, are sliced into two subviews by the user- and item-side respectively, and selectively combined into subview pairs bilaterally based on the characteristics of inter-node relation structure. Furthermore, to align the distribution of user and item embeddings after aggregation, a dispersing loss is leveraged to adjust the mutual distance between all embeddings for maintaining learning ability. Comprehensive experiments on two public datasets have proved the superiority of BusGCL in comparison to various recommendation methods. Other models can simply utilize our bilateral slicing contrastive learning to enhance recommending performance without incurring extra expenses.

Existing Scene Text Recognition (STR) methods typically use a language model to optimize the joint probability of the 1D character sequence predicted by a visual recognition (VR) model, which ignore the 2D spatial context of visual semantics within and between character instances, making them not generalize well to arbitrary shape scene text. To address this issue, we make the first attempt to perform textual reasoning based on visual semantics in this paper. Technically, given the character segmentation maps predicted by a VR model, we construct a subgraph for each instance, where nodes represent the pixels in it and edges are added between nodes based on their spatial similarity. Then, these subgraphs are sequentially connected by their root nodes and merged into a complete graph. Based on this graph, we devise a graph convolutional network for textual reasoning (GTR) by supervising it with a cross-entropy loss. GTR can be easily plugged in representative STR models to improve their performance owing to better textual reasoning. Specifically, we construct our model, namely S-GTR, by paralleling GTR to the language model in a segmentation-based STR baseline, which can effectively exploit the visual-linguistic complementarity via mutual learning. S-GTR sets new state-of-the-art on six challenging STR benchmarks and generalizes well to multi-linguistic datasets. Code is available at https://github.com/adeline-cs/GTR.