Self-Supervised User Embedding Alignment for Cross-Domain Recommendations Via Multi-LLM Co-training

Cross-domain recommendation is the central problem of personalized systems, especially in those cases where user prediction needs to be promoted across different kinds of content (e.g., music to books, or videos to articles). Typical collaborative filtering models cannot work well in this environment because user behavior representations are usually inconsistent in different domains and most existing neural models have a problem of overfitting on individual domains and lack generalization. This paper presents a new self-supervised dual-LLM model of getting user embeddings to transfer across domains through multi-view representations. In particular, two domain-specificities LLMs are trained simultaneously, sharing layers of user embedding. Such embeddings are considered to be cross-modal semantic anchors, which are learned through a joint contrastive loss (where semantically similar user profiles in a domain are brought closer to each other) and a reconstruction loss (where intra-domain similarity should be maintained). This combined goal allows this model to map high-dimensional user embeddings in domain $A$ and $B$ into a common latent space so that in the target domains where user history may be sparse. It takes an advantage of both Masked Language Modelling (MLM), and next-item prediction tasks to retain modality-specific learning in both LLMs, and also use contrastive learning enforcement of alignment using positive pairs sampled based on implicit behavioral cues. On cross-domain benchmarks, the experimental results indicate an improved top-k accuracy, embedding similarity (via CKA and t-SNE distance) and across transfer tasks compared to current image matching approaches CDAN, DGRec and CoNet.