SemaCDR: LLM-Powered Transferable Semantics for Cross-Domain Sequential Recommendation

A team of researchers, including Chunxu Zhang and Shanqiang Huang, has introduced SemaCDR, a novel framework designed to tackle the persistent challenges of data sparsity and cold-start users in recommendation systems. Traditional cross-domain recommendation (CDR) methods often fail because they rely on features or identifiers that don't translate well between different platforms, like trying to recommend a movie based on someone's book purchases. SemaCDR overcomes this by leveraging the semantic understanding power of large language models (LLMs) to construct a unified, transferable semantic space for items across domains.

The core innovation of SemaCDR is its two-stage process. First, it systematically uses an LLM to generate both domain-specific and domain-agnostic semantic descriptions for items. These LLM-powered features are then integrated with traditional content features and aligned using contrastive learning, a technique that teaches the model which items are semantically similar regardless of their original domain. Second, an adaptive fusion mechanism synthesizes user interaction sequences from source, target, and mixed domains to create a robust, transferable representation of user preferences.

Extensive testing on real-world datasets demonstrates that SemaCDR consistently outperforms existing state-of-the-art baselines. The framework proves particularly effective at capturing coherent patterns within a single domain while successfully facilitating the transfer of that knowledge to a new, data-poor domain. This represents a significant step beyond models that simply share user IDs or shallow embeddings, moving towards a deeper, semantics-based understanding of user intent that can bridge disparate platforms.