Can discrete codes transfer better than text embeddings?

Pre-trained-language-model-based transferable recommenders use the paradigm "text → representation": encode item title and description with a PLM, use the encoding as the item embedding. This works for cross-domain transfer because language is universal — but it has two failure modes. First, the recommender becomes too dependent on text similarity rather than interaction sequences, so it tends to recommend items with similar descriptions even when sequential evidence says otherwise. Second, text encodings from different domains live in different subspaces, so the domain gap survives the encoding step.

VQ-Rec inserts an intermediate representation: "text → code → representation." Item text is mapped via Optimized Product Quantization to a vector of discrete indices (the item code), and the code looks up embeddings that get aggregated. Text influence is mediated through the code rather than direct.

Two consequences. First, the discrete code distributes items more uniformly across the code space, making them more distinguishable than continuous text encodings tend to be. Second, the code-to-embedding mapping is parameter-efficient and can be tuned per downstream domain, while the text-to-code mapping stays fixed. Adapting to a new domain becomes a small fine-tune of an embedding table rather than retraining an encoder. The general principle: when transfer fails, look for the place where two representations are too tightly coupled, and insert a discrete intermediate that breaks the coupling.

Source: Recommenders Architectures