Recommender systems frequently encounter the "cold-start" problem: how can a platform suggest content to a user who has no interaction history? A powerful solution lies in Cross-Domain Recommendation (CDR), which leverages a user’s rich history in a source domain (e.g., movies) to predict their preferences in a sparse target domain (e.g., books).
The core challenge of CDR is "translation." If we represent a user as a vector in a latent space, how do we move that vector from the movie space to the book space? While early methods relied on linear transformations, a growing body of research demonstrates that human preferences are far too complex for simple projections. Instead, the field has shifted toward non-linear latent space mapping, using deep neural networks to capture the intricate ways tastes evolve across different contexts.
The Failure of Linear Proportionality
Traditional cross-domain techniques often assumed that user preferences transfer linearly. In this view, if you like Action Movie A twice as much as Movie B, you might like Action Book A twice as much as Book B. However, as Xin et al. (2015) argue, human behavior is rarely so proportional. They point out that preferences are often multi-modal: for instance, both extreme fans and extreme critics of a specific movie genre might share a high interest in a particular book category. A linear matrix cannot capture this "U-shaped" or complex relationship.
Similarly, Shen et al. (2017) note that the latent relationships between distinct domains are inherently complex. Linear assumptions fail to represent the nuanced shifts that occur when a user moves between platforms. To address this, the authors proposed the EMCDR (Embedding and Mapping for Cross-Domain Recommendation) framework, which formalizes a two-stage process: first, learn latent embeddings for users and items in each domain independently; second, learn a non-linear mapping function—typically a multi-layer perceptron (MLP)—to bridge them.
Bridging Heterogeneous Latent Spaces
The necessity for non-linearity becomes even more apparent when the domains are heterogeneous, meaning they have different data structures or semantic granularities.
- Social to E-commerce: Wang et al. (2017) demonstrated that deep multi-layer architectures are essential for bridging the gap between connection-rich social networks and item-rich information domains. The depth of the network allows the model to reconcile the "heterogeneity gap" between social behaviors and purchasing patterns.
- Semantic Granularity: Yu et al. (2019) highlight the difficulty of mapping fine-grained data (like specific tags on Twitter) to broad categories (like video genres). They argue that this mapping is fundamentally non-linear, requiring deep learning to capture how specific, narrow interests aggregate into broader preferences.
- Interest Drift: When users change environments—such as traveling to a new city—their preferences undergo "interest drift." Xin et al. (2021) show that simple projections of "home-town" preferences fail to capture how travel intentions (business vs. leisure) transform behavior. Non-linear mapping functions are required to model this contextual evolution.
Architectures for Preference Translation
To perform these translations, researchers have moved beyond basic MLPs to more specialized non-linear architectures:
- Kernel Methods: Xin et al. (2015) utilized Radial Basis Function (RBF) kernels to map preferences, allowing the model to handle the complex, multi-modal correlations mentioned earlier.
- Gradient Boosting: Wang et al. (2018) introduced the use of Gradient Boosting Trees (GBT) to learn the mapping. GBTs are particularly effective at capturing higher-order dependencies and non-proportional relationships between latent features.
- Autoencoders: In the context of transferring internal representations between models, Oozeer et al. (2025) found that non-linear autoencoders are superior to linear maps for translating activation spaces. This suggests that the "universality" of latent representations across different architectures is best unlocked through non-linear bridges.
- Encoder-Decoders: For news recommendation, Hu et al. (2021) employed non-linear encoder-decoder structures to translate user interests across vertical news platforms with different vocabularies.
Why Non-linearity Works
The shift toward non-linearity, as summarized by Zhu et al. (2018), provides the flexibility needed to model the "intricacy" of real-world interactions. While linear models assume a direct, one-to-one relationship between features, non-linear models like those in Yao et al. (2018) can map sparse, high-dimensional interaction history into dense latent spaces by identifying complex behavioral patterns.
By treating the relationship between domains as a learnable, deep transformation rather than a fixed geometric projection, these models allow recommendation systems to "understand" that a user's love for a specific director might manifest not just as a preference for similar movies, but as a specific, non-obvious interest in a particular style of historical biography or travel destination.
Key Foundational Work
Cross-Domain Collaborative Filtering with Review Text (2015) — contributed 4 concept elements to this direction.
Cross-Domain Recommendation: An Embedding and Mapping Approach (2017) — contributed 3 concept elements to this direction.
Item Silk Road: Recommending Items from Information Domains to Social Users (2017) — contributed 1 concept elements to this direction.