AI RESEARCH
Anchored Alignment: Preventing Positional Collapse in Multimodal Recommender Systems
arXiv CS.LG
•
ArXi:2603.12726v1 Announce Type: cross 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.