Cross-Domain Recommendation

The system learns a unified user embedding that captures preferences spanning multiple domains. A user's behavior in the source domain (e.g., book purchases) generates a latent vector that directly initializes their profile in the target domain (e.g., movie recommendations).

• Overlapping users serve as the bridge for training the shared representation space
• Multi-task learning jointly optimizes objectives across both domains simultaneously
• The shared layer forces the model to extract domain-invariant preference signals

When user overlap is sparse, the system learns a transformation function that maps user or item embeddings from the source latent space to the target latent space. This approach does not require the same users to exist in both domains.

• Linear mapping learns a transformation matrix between embedding spaces
• Non-linear neural mappers capture complex cross-domain relationships
• Training requires only a small set of known correspondences between domains
• The mapping generalizes to cold-start users who only exist in the source domain

This technique simultaneously factorizes the user-item interaction matrices of multiple domains while sharing latent factors across them. The shared factors capture cross-domain preference patterns.

• A shared user factor matrix links domains, while item factors remain domain-specific
• The factorization objective includes a cross-domain regularization term
• Handles scenarios where user sets partially overlap across domains
• Item metadata from both domains can be incorporated as side information constraints

A cross-domain knowledge graph connects entities from both domains through semantic relationships, enabling reasoning over paths that link a user's source-domain interests to target-domain items.

• Entities like authors, genres, or brands serve as bridge nodes between domains
• Graph embedding techniques (e.g., TransE, RotatE) encode multi-hop relational paths
• A user who likes a specific author in the book domain can receive film recommendations directed by that same individual
• Path-based reasoning provides explainable cross-domain recommendations

The system transfers temporal consumption patterns learned in the source domain to predict the next interaction in the target domain. A user's sequential behavior—such as browsing, adding to cart, and purchasing—exhibits transferable patterns.

• Recurrent neural networks or transformers model cross-domain session sequences
• The model learns that a purchase in one domain often precedes a related search in another
• Session-level transfer captures short-term intent shifts across domains
• Particularly effective for cross-domain next-basket recommendation in e-commerce ecosystems

A domain classifier is trained adversarially to ensure the shared user representation contains no domain-specific information, forcing the encoder to extract only transferable preference features.

• A gradient reversal layer flips gradients during backpropagation to confuse the domain discriminator
• The encoder learns to produce representations that are simultaneously predictive of preferences and invariant to domain origin
• This technique prevents negative transfer where domain-specific noise degrades target-domain performance
• Results in a domain-agnostic preference space that generalizes to entirely new domains

The foundational machine learning paradigm that enables cross-domain recommendation. A model trained on a source domain with abundant data is repurposed as the starting point for a target domain with sparse interactions.

• Pre-training: Model learns general user preference patterns from a dense domain (e.g., movie ratings)
• Fine-tuning: Adapted to a sparse domain (e.g., book recommendations) with minimal new data
• Domain divergence risk: Negative transfer occurs when source and target domains are too dissimilar

Dense vector representations learned from a large, general-purpose dataset and reused across domains. These embeddings encode semantic relationships that transcend specific item catalogs.

• A user embedding trained on e-commerce browsing can initialize a news recommendation profile
• Zero-shot capability: Embeddings provide meaningful similarity scores even for items never seen in the target domain
• Common sources include BERT, GPT, and domain-specific models like Prod2Vec

A paradigm where a model learns how to learn across many tasks, enabling rapid adaptation to new domains with few examples. Often formalized as Model-Agnostic Meta-Learning (MAML).

• Trains on a distribution of related recommendation tasks
• Finds an initialization that can adapt to a new domain in just a few gradient steps
• Critical for few-shot user modeling when a user has only 1-3 interactions in the target domain

Translates entities and relationships from a structured knowledge graph into low-dimensional vectors. These embeddings bridge domains by capturing semantic connections between items across catalogs.

• A movie entity and a book entity can be linked through shared attributes like genre, author, or theme
• Techniques like TransE and RotatE model relational patterns
• Enables cold-start recommendations by traversing cross-domain entity relationships

A subfield of transfer learning specifically addressing the shift in data distributions between source and target domains. Techniques align feature spaces to make knowledge transfer possible.

• Adversarial domain adaptation: Uses a domain classifier to learn domain-invariant representations
• Maximum Mean Discrepancy (MMD): Minimizes the statistical distance between source and target feature distributions
• Essential when user behavior patterns differ fundamentally between domains (e.g., music vs. grocery shopping)

Extends traditional matrix factorization to jointly factorize interaction matrices from multiple domains, sharing latent factors across them.

• Collective Matrix Factorization (CMF): Simultaneously decomposes multiple matrices sharing user or item factors
• Cross-domain tri-factorization: Models user-domain-item interactions as a three-way tensor
• Enables knowledge transfer even when user sets only partially overlap between domains