Side Information

Effective side information must be strongly correlated with the target behavior. Irrelevant attributes introduce noise that degrades model performance rather than improving it.

• Demographic signal: A user's age and location are highly predictive for mortgage offers but nearly useless for music taste.
• Categorical signal: An item's brand and sub-category are strong predictors for fashion, while color alone is weak.
• Validation: Use mutual information scores to quantify the statistical dependency between a side feature and the target variable before integration.

Side information is only useful if it exists before or at the exact moment of the cold start event. Data that arrives late cannot bootstrap the initial experience.

• User registration: Attributes like sign-up source, device type, and referral code are available at t=0.
• Item ingestion: Catalog metadata such as SKU, price, and description are available the moment an item enters the system.
• Architecture requirement: A feature store must serve these pre-computed attributes with sub-millisecond latency to the online inference endpoint.

Raw side information must be transformed into a dense vector representation that a neural network can consume. The quality of this embedding directly impacts cold start accuracy.

• Categorical encoding: High-cardinality fields like product_category are mapped to learned embeddings via an entity embedding layer.
• Textual encoding: Free-text descriptions are passed through a pre-trained transformer like Sentence-BERT to generate a 768-dimensional semantic vector.
• Numerical normalization: Continuous features like price or age are standardized to zero mean and unit variance to stabilize gradient descent.

Sparse, low-information features fail to create meaningful similarity links. Rich, semantically dense side information enables precise nearest-neighbor matching in the latent space.

• Knowledge graph integration: Linking an item to a node in a knowledge graph provides multi-hop relational context (e.g., this film is directed by a director known for neo-noir).
• Multi-modal enrichment: A product image processed by a Vision Transformer (ViT) provides visual side information that text alone cannot capture.
• Contrastive pre-training: Side information encoders are often pre-trained using a contrastive loss to pull similar entities close together in the embedding space before the cold start occurs.

Side information that changes rapidly or is inconsistently labeled creates a moving target that undermines the model's ability to learn stable initial representations.

• Static attributes: A user's birth year is stable; their inferred mood is not. Prioritize stable attributes for cold start bootstrapping.
• Taxonomy governance: Product categories must follow a strictly governed taxonomy. A sudden renaming of "Sneakers" to "Athletic Footwear" breaks the mapping for all new items.
• Schema validation: Implement data contracts that reject incoming items with missing or malformed side information fields before they reach the model.

The most powerful side information is domain-agnostic, allowing a user profile built in one vertical to bootstrap recommendations in a completely unrelated vertical.

• Natural language as a universal schema: A user's self-described interests in a free-text bio can be embedded and applied to recommend books, movies, or travel destinations.
• Pre-trained foundation embeddings: Vectors from models like CLIP or BERT encode general world knowledge that transfers across siloed product catalogs.
• Cross-domain mapping: A user's affinity for "minimalist design" in furniture can be transferred to recommend "minimalist" electronics or fashion items via a shared semantic attribute space.