Why do recommendation models fail when new users arrive?

Most published recommendation methods assume the algorithm sees, at training time, all the users and items that will appear at test time. This is the transductive setting: matrix factorization implicitly assumes test users equal training users. The convenience of the assumption is enormous because it lets the model learn one embedding per entity. The cost is that on real platforms — where new users sign up, new items get uploaded, and interactions happen continuously — the model is structurally unable to score the entities it most needs to handle.

The survey by Sun, Lee, and Goldberg formalizes the gap. Inductive learning works on unseen user-item contexts at test time. User-inductive: known item set, unknown users. Item-inductive: known users, unknown items. Cold-start is a separate axis (insufficient information about an entity) orthogonal to whether the model can structurally accept the entity at all.

Feature-based approaches are one route to inductive recommendation but suffer when features are unavailable, can't be accessed for privacy reasons, or can't represent latent aspects (e.g., aesthetic taste in art domains where features are unclear). Aggregation-based methods work but face a "curse of directionality" — they have to choose to represent items through users or vice versa because the forward pass is one-directional. The open question is whether models can be designed with cyclic aggregation between users and items for full inductivity. The deeper claim: latent factors plus gradient backprop pose fundamental obstacles to inductive learning, and future work should explore non-parametric and tree-based families.

Source: Recommenders Architectures