Anchored Alignment: Preventing Positional Collapse in Multimodal Recommender Systems

Multimodal recommender systems (MMRS) have become a hot research area in recent years. Traditional recommender systems primarily rely on user-item interaction data, but as data becomes more diverse, single-modality data is no longer sufficient to meet user needs. Early MMRS, such as VBPR and FREEDOM, integrated visual and textual features into ID-based recommendation frameworks to partially alleviate data sparsity and cold-start issues. However, these methods typically treat multimodal signals as auxiliary features, simply integrating them through fusion mechanisms, leading to insufficient cross-modal alignment. To overcome this limitation, recent alignment-based methods, such as DA-MRS and AlignRec, explicitly project all modalities into a unified latent space to achieve better cross-modal consistency.

Although alignment-based MMRS improve cross-modal consistency to some extent, they introduce a fundamental trade-off: converging all modalities into a single space reduces modality-specific expressiveness. This trade-off leads to two main challenges: positional collapse of modality representations and overdominance of interaction signals. Positional collapse refers to the compression of embeddings from different modalities into nearly identical positions, reducing semantic diversity and diminishing modality-specific characteristics. Overdominance of interaction signals occurs when interaction-driven objectives heavily bias the final item embeddings towards ID-based interaction patterns, suppressing multimodal semantics.

AnchorRec's core innovation lies in its anchor-based alignment strategy. First, it avoids the loss of modality-specific expressiveness by performing indirect alignment in a lightweight projection domain. Second, AnchorRec designs a fused multimodal embedding as an anchor, providing a stable semantic reference that guides the projected ID, text, and vision representations toward semantic agreement. Unlike existing methods, AnchorRec does not force all modalities to overlap within a single latent space but achieves alignment in the projection domain, preserving modality-specific structures while avoiding positional collapse.

AnchorRec's methodology includes the following key steps:

• �� Modality Encoders: Use pretrained modality encoders to extract modality-specific features for each item, including textual, visual, and multimodal fusion features.

• �� Collaborative Refinement: Inject interaction-driven information into item modality embeddings and construct user-side modality preferences to address the lack of modality features on the user side.

• �� Anchor-based Projection: Map item-side modality embeddings into the projection domain and achieve alignment using anchor-based alignment loss.

• �� Representation Fusion: Fuse signals from ID, multimodal, textual, and visual embeddings to obtain the final item representation.

The experimental design includes evaluations on four real-world datasets: Baby, Sports, Office, and Video Games. Each dataset contains user-item interactions along with textual descriptions and images for each item. Experiments use Recall@20 and NDCG@20 as the main evaluation metrics and compare against various baseline methods, including VBPR, LATTICE, FREEDOM, LGMRec, SMORE, BM3, DA-MRS, and AlignRec. Results show that AnchorRec performs well across multiple datasets and metrics, particularly on the Baby dataset where Recall@20 reached 0.1007.

Experimental results indicate that AnchorRec achieves competitive top-N recommendation accuracy, especially on the Baby dataset where Recall@20 reached 0.1007. Qualitative analyses demonstrate improved multimodal expressiveness and coherence, particularly in the integration of visual and textual features, significantly outperforming existing methods. Ablation studies reveal that AnchorRec's anchor-based alignment strategy effectively reduces ID signal dominance, enhancing the balance of multimodal signals.

AnchorRec has broad application potential in multimodal recommender systems. It can be directly applied to e-commerce platforms to improve recommendation accuracy and personalization by integrating users' multimodal preferences. Additionally, AnchorRec can be used in content recommendation systems, such as news and video recommendations, to provide a richer recommendation experience by combining textual and visual features.

Despite AnchorRec's strong performance in multimodal recommender systems, it may perform poorly when dealing with users lacking modality features. Additionally, due to the complexity of the anchor-based alignment strategy, AnchorRec may not be as computationally efficient as some simpler fusion methods. In certain specific application scenarios, AnchorRec may require additional adjustments and optimizations for specific modality features. Future research directions include further optimizing AnchorRec's computational efficiency, exploring its performance on more diverse datasets, and applying it to real-time recommendation systems.