The "cold-start" problem is a fundamental hurdle in recommendation systems: how do you suggest items to a user you have never seen before? While a new user on a book platform might have zero reading history, they often have rich histories elsewhere—perhaps on a movie streaming service or a news aggregator.
The research direction of Cross-Domain Recommendation (CDR) aims to bridge these silos. The core intuition is that user preferences are not isolated; a person’s taste in movies likely correlates with their taste in books. However, recent research has shifted away from simple correlations toward a more sophisticated mechanism: non-linear latent mapping.
The Failure of Linear Assumptions
Early attempts at cross-domain transfer relied on linear transformations—essentially using a single matrix to "rotate" or "scale" a user's preference vector from one domain to another. This assumes a proportional relationship: if you like action movies twice as much, you’ll like action books twice as much.
As Xin Xin and colleagues argue in Cross-Domain Collaborative Filtering with Review Text, human preferences are rarely that simple. They are often multi-modal or "non-proportional." For example, both extreme fans and extreme critics of a specific movie genre might share a high interest in a related book category, a pattern that a linear model fundamentally cannot capture. To solve this, Xin et al. introduced non-linear mapping functions using Radial Basis Function (RBF) kernels, allowing the system to model these intricate, "curved" relationships between interests.
The Embedding and Mapping Paradigm
The modern standard for this approach is the EMCDR (Embedding and Mapping for Cross-Domain Recommendation) framework, popularized by Huawei Shen, Tong Man, and their team in Cross-Domain Recommendation: An Embedding and Mapping Approach.
The process follows two distinct stages:
- Latent Space Construction: Independent embeddings are learned for users and items in both the source and target domains using standard collaborative filtering.
- Non-linear Mapping: A Multi-Layer Perceptron (MLP) is trained to "translate" the user's latent vector from the source space to the target space.
By using an MLP, the model can capture the complex, non-linear transformations required to project a user’s "movie identity" into a "book identity." This flexibility is essential because, as Feng Zhu notes in A Deep Framework for Cross-Domain and Cross-System Recommendations, user behavior patterns are rarely directly proportional across different platforms.
Deep Fusion of Content and Context
To make these mappings even more robust, researchers have begun incorporating auxiliary data like review text and item descriptions. In the RC-DFM model, Wenjing Fu and Senzhang Wang use Stacked Denoising Autoencoders to extract features from reviews before applying the non-linear MLP mapping. This "deep fusion" ensures that the mapping isn't just based on sparse rating matrices but on the actual semantic content of user preferences (Deeply Fusing Reviews and Contents for Cold Start Users).
Handling Interest Drift
The power of non-linear mapping is most evident when user interests "drift" or evolve during a transition. In TrNews, Guangneng Hu and Qiang Yang demonstrate that an interest in general news manifests as very specific, niche preferences when moving to a new platform—a transformation that requires a non-linear encoder-decoder architecture to approximate.
Similarly, Haoran Xin addresses the "out-of-town" recommendation problem in TRAINOR. When a user travels to a new city, their behavior changes; they aren't looking for the same things they do at home. By using non-linear mapping to translate "home-town" preference vectors into "travel" contexts, the system can account for this behavioral evolution, recognizing that the relationship between a user's routine and their travel intentions is complex and context-dependent.