Matrix Factorization

The core mechanism that uncovers hidden dimensions of preference from observed interactions. Instead of relying on explicit attributes, the model learns abstract features—such as a user's affinity for 'gritty sci-fi' or 'Scandinavian minimalism'—directly from rating patterns. Each user and item is represented by a vector of k latent factors, where the dot product between a user vector and an item vector predicts the user's preference for that item. This dimensionality reduction forces the model to generalize, capturing the underlying structure that explains observed behavior rather than memorizing it.

Formally, the user-item interaction matrix R (of size m × n) is approximated by the product of two lower-rank matrices: R ≈ P × Qᵀ. Here, P is the m × k user factor matrix and Q is the n × k item factor matrix. The optimization objective is typically to minimize the regularized squared error between observed and predicted ratings:

• Objective: min Σ (rᵤᵢ - pᵤ · qᵢ)² + λ(||pᵤ||² + ||qᵢ||²)
• Regularization parameter λ prevents overfitting on sparse data
• Stochastic Gradient Descent (SGD) or Alternating Least Squares (ALS) are used to iteratively learn the factor vectors

Standard collaborative filtering fails for new items with no interactions. Matrix Factorization addresses the item cold start by incorporating side information through feature mapping. Instead of learning an item's latent vector solely from interactions, the vector is constructed from its content attributes:

• A new item's feature vector x (e.g., brand, category, price) is transformed via a learned mapping function: q_new = f(x)
• This allows immediate placement in the latent space before any user interacts with the item
• The technique bridges content-based filtering and collaborative filtering into a unified hybrid model

Raw ratings are confounded by systematic biases that obscure true user-item affinity. Matrix Factorization models explicitly capture these as learned parameters:

• Global bias (μ): The average rating across all users and items
• User bias (bᵤ): Some users consistently rate higher or lower than average
• Item bias (bᵢ): Some items are universally perceived as better or worse
• Full prediction: r̂ᵤᵢ = μ + bᵤ + bᵢ + pᵤ · qᵢ By isolating these biases, the latent factors focus purely on modeling the interaction effect—the specific affinity between a user and an item beyond baseline tendencies.

Beyond explicit ratings, modern Matrix Factorization variants incorporate implicit feedback signals—clicks, views, purchase history, and dwell time—which are far more abundant. The model adds a second set of item factors to capture which items a user has interacted with, regardless of rating:

• SVD++ extends the basic model by adding a term: Σ (yⱼ / √|N(u)|) where N(u) is the set of items the user has implicitly interacted with
• This provides a strong signal for user cold start scenarios, as a new user's first few clicks immediately influence their latent representation
• The approach models both preference (how much they liked it) and confidence (that they interacted at all)

User preferences and item perception drift over time. TimeSVD++ extends Matrix Factorization by making factors time-dependent:

• User factors evolve as a function of time: pᵤ(t) = pᵤ + αᵤ · dev(t) + pᵤ,ₜ
• Item biases shift as products age or gain popularity: bᵢ(t) = bᵢ + bᵢ,ₜ
• This captures seasonal effects, changing tastes, and the lifecycle of products
• Critical for dynamic retail hyper-personalization where a user's intent during a holiday sale differs fundamentally from their browsing behavior in off-peak periods

The process of incorporating auxiliary metadata—such as item brand, category, price, or user demographics—directly into the factorization framework. Instead of relying solely on interaction history, the model learns a mapping function from these features to the latent space. For a new item, its metadata is passed through this learned mapping to generate an initial embedding vector, enabling immediate similarity comparisons with existing items before any user interactions occur.

An extension of standard matrix factorization that models all feature interactions, not just user-item pairs. Factorization Machines treat each data point as a sparse feature vector and learn a latent factor for every feature, not every user or item. This allows the model to compute predictions for completely new users or items by summing the interactions of their observed features, making it inherently capable of handling cold starts without retraining.

A multi-view learning technique that simultaneously factorizes multiple related matrices—such as the user-item interaction matrix alongside an item-attribute co-occurrence matrix. By sharing the item latent factors across both decompositions, the model constrains the learned embeddings to be consistent with both interaction patterns and content metadata. A new item's factors are initialized purely from its attributes, grounding cold-start recommendations in its intrinsic properties.

A hybrid approach where a deep neural network is trained to predict the latent factor vector of an item directly from its content features. The training process uses the embeddings of warm-start items as supervised targets. Once trained, the network acts as a universal embedding projector: feed in a new item's features, and it outputs a high-quality latent vector that slots directly into the existing collaborative filtering model without requiring any gradient updates to the original factorization.

A probabilistic formulation that places informative priors on the latent factor distributions, tied to observable side information. Rather than starting with a zero-mean Gaussian prior for new items, the prior mean is set based on the item's category or brand. This shrinks the initial estimate toward a sensible group-level baseline. As interactions accumulate, the posterior distribution shifts from the prior to the data-driven estimate, providing a statistically principled transition from cold to warm start.

An extension that models how latent factors evolve over time, capturing shifts in user taste or item relevance. For cold-start items, this framework incorporates a lifecycle prior that predicts how a new item's embedding will change in its first hours or days based on patterns observed in similar past items. This allows the system to anticipate the typical adoption curve and adjust recommendations proactively rather than waiting for sufficient data to accumulate.