Organizations: Harbin Institute of Technology(Shenzhen) Shenzhen, China · Shenzhen Key Laboratory of Internet Information Collaboration Shenzhen, China · Shenzhen Research Institute, City University of Hong Kong Shenzhen, China · Nanyang Technological University Singapore, Singapore
Recommender systems often suffer from over-concentrated results when optimized for accuracy, limiting recommendation diversity. While diversity has become an important objective, existing methods still face fundamental limitations from both semantic and structural perspectives. On the one hand, multimodal recommender systems provide rich semantic signals, yet are often treated as auxiliary to ID embeddings, leading to ineffective semantic utilization for diversity. On the other hand, the accuracy--diversity trade-off persists during training, where accuracy gains often come at the cost of diversity. We show that this phenomenon is fundamentally governed by both the structural properties of learned representations and the dynamics of optimization. To address these challenges, we propose MOSS, a unified framework for multimodal decoupling through semantic and structural shaping. Specifically, we first introduce an MLLM-based Decoupled Feature Extraction module to disentangle multimodal content into informative representations. To enhance semantic utilization, we design a Category-aware Embedding Selection module that extracts compact, category-aware semantic features. To address structural limitations, we propose two complementary modules: Alternating Dual-Graph Propagation, which alleviates representation collapse by integrating collaborative and multimodal propagation, and Entropy-based Diversity Regularization, which mitigates optimization bias by promoting balanced category exposure. Extensive experiments on real-world datasets demonstrate that MOSS consistently achieves a superior accuracy--diversity trade-off over strong baselines.
Figures & tables
Figure 1. Training dynamics and the accuracy–diversity trade-off of LightGCN on Anime, with Recall@10 for accuracy and Coverage@10 for diversity. The left figure shows Recall@10 and Coverage@10 over the first 50 epochs, and the right figure shows the corresponding Coverage@10–Recall@10 trajectory, colored by epoch and labeled every five epochs.
Features
Norm. Eff.Rank
R@10
C@10
ID embedding
0.99
0.1924
0.4074
Visual embedding
0.65
0.1710
0.4032
Textual embedding
0.68
0.1698
0.4096
Table 1. Effect of different item initialization strategies on representation quality and recommendation performance on Anime dataset.
Figure 2. Category-wise density visualization of item embeddings for four representative genres on Anime dataset. Embeddings are projected into 2D by t-SNE and normalized onto the unit circle, with filled contours represent density.
Figure 3. Category-aware Embedding Selection.
Figure 4. An illustration of the proposed MOSS model. The MLLM-based Decoupled Feature Extraction module disentangles multimodal content into informative representations. The Category-aware Embedding Selection module selects compact, category-aware features to enhance semantic utilization. The Alternating Dual-Graph Propagation module alleviates dimensional collapse via collaborative and multimodal propagation. The Entropy-based Diversity Regularization module promotes balanced category exposure while preserving accuracy.
Figure 5. Comparison of the top- 50 normalized singular value of two two-layer propagation operators on Anime. The curves show the normalized singular value ratios σj/σ1 for Q2UI and QUI+II , sorted by singular value magnitude.
Table 2. Statistics of datasets.
Dataset
Metric
Accuracy-based Methods
Diversity-based Methods
Our Method
LightGCN
SimGCL
Freedom
DGCN
DGRec
DRGame
DivGCL
MOSS
MOSS acc
Anime
Recall@10
0.1924
0.2100
0.1985
0.1106
0.2024
0.0569
0.2132
0.1860
0.2211
NDCG@10
0.2124
0.2248
0.2014
0.1169
0.1868
0.0975
0.2329
0.1756
0.2374
Coverage@10
0.4074
0.4387
0.4345
0.4890
0.4428
0.4279
0.4197
0.5536
0.4417
Entropy@10
5.0054
5.7738
5.3682
6.1741
5.5891
5.6576
5.1921
6.9645
6.0342
GM@10
0.2800
0.3035
0.2937
0.2326
0.2994
0.1560
0.2991
0.3209
0.3125
Table 3. Comparison of models on the Anime and Game datasets.
Anime
Game
Model
Recall@10
NDCG@10
Coverage@10
Entropy@10
GM@10
Recall@10
NDCG@10
Coverage@10
Entropy@10
GM@10
w/oEBDR
0.2172
0.2306
0.4577
6.0258
0.3153
0.1092
0.0607
0.3642
2.9347
0.1994
w/oADGP
0.1730
0.1619
0.5498
6.5569
0.3084
0.0863
0.0384
0.6127
3.8642
0.2299
w/oCAES
0.1722
0.1591
0.5187
6.5487
0.2989
0.0852
0.0372
0.5932
3.8533
0.2248
MOSS
0.1860
0.1756
0.5536
6.9645
0.3209
0.0907
0.0391
0.6348
3.9764
0.2400
Table 4. Ablation study on the Anime and Game datasets.
Figure 6. Comparison of the geometric mean (GM) between MOSS and baseline methods.
Multimodal Graph Neural Networks have become standard for recommendation by augmenting sparse interaction data with content features. Yet current architectures face two bottlenecks: structural rigidity, from a reliance on static precomputed similarity graphs that cannot adapt to evolving preferences; and semantic fragility, where noisy modality signals are indiscriminately fused, distorting the collaborative signal. We propose MURAL (Multimodal Uncertainty-aware Recommendation via Adaptive edge Learning), a unified framework that shifts multimodal recommendation from fixed structural augmentation to dynamic topology discovery. To address structural rigidity, an Adaptive Edge Learner combines a differentiable retrieval-augmented strategy with an approximate nearest neighbor search to discover latent item-item correlations that are both semantically adaptive and computationally scalable (O(NlogN)). To address semantic fragility, an Uncertainty-Aware Fusion module models the aleatoric uncertainty of heterogeneous modalities, dynamically down-weighting unreliable features while prioritizing high-confidence signals as a defense against cross-modal noise. We further employ a contrastive teacher-student alignment that anchors modality-specific representations to stable behavioral signals, ensuring optimization stability without gradient leakage. Experiments on large-scale benchmarks including TikTok and Amazon show that MURAL significantly surpasses both structural and generative state-of-the-art baselines, achieving superior accuracy while offering interpretability through domain-specific modality dominance and robustness under extreme data corruption.
Ahmad Mousavi, Majid Alikhani, Yeon-Chang Lee +2
Department of Mathematics and Statistics American University · Independent Researcher Toronto, Ontario, Canada · Independent Researcher +4
Multimodal recommender systems (MMRS) leverage images, text, and interaction signals to enrich item representations. However, recent alignment based MMRSs that enforce a unified embedding space often blur modality specific structures and exacerbate ID dominance. Therefore, we propose AnchorRec, a multimodal recommendation framework that performs indirect, anchor based alignment in a lightweight projection domain. By decoupling alignment from representation learning, AnchorRec preserves each modality's native structure while maintaining cross modal consistency and avoiding positional collapse. Experiments on four Amazon datasets show that AnchorRec achieves competitive top N recommendation accuracy, while qualitative analyses demonstrate improved multimodal expressiveness and coherence. The codebase of AnchorRec is available at https://github.com/hun9008/AnchorRec.
Yonghun Jeong, David Yoon Suk Kang, Yeon-Chang Lee
UNIST, Ulsan, Korea · Chungbuk National University, Cheongju, Korea
Multimodal recommender systems leverage diverse data sources, such as user interactions, content features, and contextual information, to address challenges like cold-start and data sparsity. However, existing methods often suffer from one or more key limitations: processing different modalities in isolation, requiring complete multimodal data for each interaction during training, or independent learning of user and item representations. These factors contribute to increased complexity and potential misalignment between user and item embeddings. To address these challenges, we propose DReX, a unified multimodal recommendation framework that incrementally refines user and item representations by leveraging interaction-level features from multimodal feedback. Our model employs gated recurrent units to selectively integrate these fine-grained features into global representations. This incremental update mechanism provides three key advantages: (1) simultaneous modeling of both nuanced interaction details and broader preference patterns, (2) eliminates the need for separate user and item feature extraction processes, leading to enhanced alignment in their learned representation, and (3) inherent robustness to varying or missing modalities. We evaluate the performance of the proposed approach on three real-world datasets containing reviews and ratings as interaction modalities. By considering review text as a modality, our approach automatically generates interpretable keyword profiles for both users and items, which supplement the recommendation process with interpretable preference indicators. Experiment results demonstrate that our approach outperforms state-of-the-art methods across all evaluated datasets.