Can discretizing text embeddings improve recommendation transfer?

When a sequential recommender uses pre-trained language model encodings as item representations, the binding between text and recommendation behavior becomes too tight. Two problems result: the recommender starts emphasizing text features (generating items with similar titles instead of similar interaction patterns), and text encodings from different domains live in different subspaces, so the domain gap in text directly causes a performance drop in cross-domain transfer.

VQ-Rec inserts a discretization step. Item text encodings are quantized through optimized product quantization into a vector of discrete codes (the "code"), and the actual item representation is constructed by looking up and aggregating embeddings indexed by that code. Text influences the code, the code influences the representation, but the representation is no longer a function of text — it's a function of which embedding cells the code addresses.

The benefits compound. The codes are uniformly distributed over the item set, making them highly distinguishable. The two mappings (text→code, code→embedding) are independently tunable: the lookup table can be adapted to a new domain without modifying the text encoder. And because the backbone (Transformer) is unchanged, the technique drops into existing sequential architectures. The decoupling is the point — text becomes a semantic feeder, not the representation itself.

Source: Recommenders Architectures