Cross-Domain Recommendation: Concepts, Methods, and Novel Approaches

Recommendation systems play an increasingly important role in modern online services, but single‑domain personalized recommendation suffers from data sparsity, cold‑start, and limited personalization. Cross‑domain recommendation combines data from multiple domains (e.g., books and movies) to alleviate these issues.

Traditional single‑domain approaches rely on collaborative filtering, which predicts user preferences based on historical interactions. However, sparse user‑item matrices and new users limit effectiveness.

Cross‑domain techniques address these problems by leveraging auxiliary domains. Benefits include mitigating sparsity, solving cold‑start, and enabling truly personalized recommendations.

Existing cross‑domain methods fall into three categories:

Collaborative‑filtering based: matrix merging (concatenating user‑item matrices) and collective matrix factorization (CMF) with extensions such as HeteroMF that add domain‑specific weights.

Transfer‑learning based: codebook transfer, where clustered user‑item patterns from one domain are transferred to another, though effectiveness depends on domain relatedness.

Knowledge‑based: recommendation using external knowledge (e.g., geographic location to suggest music) without direct user‑item feedback.

Two novel approaches are proposed:

3.1 Tensor‑factorization based collaborative filtering (CDTF): models the three‑way relation using a non‑regular tensor factorization that allows each domain to have a different number of items. User latent matrix U captures shared preferences, domain matrix D captures domain‑specific traits, and their interaction yields domain‑specific user features. Weight variables, optimized via a genetic algorithm, control the influence of auxiliary domains.

3.2 Bilinear Multilevel Model (BLMA): incorporates multi‑level factors such as market bias, group bias, and individual bias into a bilinear formulation similar to matrix factorization. It models hierarchical influences on user ratings, enabling robust predictions for cold‑start users. A parallel Gibbs sampling algorithm is used for efficient parameter learning.

Extensive experiments on real datasets demonstrate that both CDTF and BLMA outperform mainstream baselines.

The article concludes with an outlook on the broad applicability of cross‑domain recommendation across various scenarios, including web ranking, news feeds, advertising, and emerging domains like smart homes and healthcare.