The Geometry of the Cold-Start Problem
Collaborative filtering models operate by mapping users and items into a shared latent space based on historical interactions. Mathematically, the affinity between a user $u$ and an item $i$ is often captured by the inner product of their latent vectors, $\langle \mathbf{p}_u, \mathbf{q}_i \rangle$. When the system has abundant interaction data, these vectors are optimized to reflect nuanced preferences.
However, when a new user or item enters the system, they possess zero interaction history. In the matrix factorization framework, their corresponding row or column in the interaction matrix is entirely empty. As noted by Dongsheng Li et al. in Expected Risk Minimization for Matrix Approximation-based Recommender Systems, standard matrix factorization approaches fail here because they primarily optimize for established entities rather than generalizing to cases with zero historical data. Without interactions, the system cannot compute a meaningful gradient to position these new vectors, leaving them stranded at the origin or randomly initialized. This geometric isolation is the root of the cold-start problem.
Bridging Spaces via Alignment Mappings
If we cannot position a new item using interaction data, we must rely on its intrinsic content features (such as text descriptions, images, or categorical metadata). The challenge is that content features live in an entirely different vector space—such as a high-dimensional text embedding space or a convolutional visual feature space—than the collaborative latent space.
The core intuition of space alignment is to learn a projection function $\phi(\cdot)$ that maps content features directly into the collaborative latent space.
[Content Feature Space] ---> Mapping Function ɸ(·) ---> [Collaborative Latent Space]
(Text, Images, Metadata) (User & Item Interaction Vectors)
Shiyu Chang et al. demonstrated this mechanism in A Space Alignment Method for Cold-Start TV Show Recommendations. By learning a mapping that aligns content features with user-behavior patterns, their framework allows a brand-new item to inherit a functional position in the collaborative space immediately. Consequently, the item can be evaluated against existing user vectors without requiring a single historical interaction.
Similarly, Chenyi Lei et al. leveraged this concept in Comparative Deep Learning of Hybrid Representations for Image Recommendations by introducing a dual-network architecture. Instead of relying on absolute rating predictions, the authors trained the network on relative preference triplets (User, Liked Image, Disliked Image). This strategy forces the visual feature extraction network to align its outputs directly with the relative ranking structure of the user preference space, bypassing the need for historical engagement data on new images.
Active Preference Elicitation and Information Borrowing
For new users, the problem is inverted: we lack a preference vector $\mathbf{p}_u$. One solution is active preference elicitation—asking the user a few diagnostic questions upon arrival. However, randomly selecting items to show a new user is highly inefficient and risks user fatigue.
To solve this, models must maximize the information gain of each query by leveraging the shared structure of the user population. Edwin V. Bonilla et al. formulated this in Gaussian Process Preference Elicitation using a Bayesian framework. By placing a Gaussian Process (GP) prior over the joint space of user and item features, the system does not treat the new user as an isolated entity. Instead, it "borrows" information from existing users.
When the new user provides a handful of pairwise comparisons, the GP propagates this feedback through the learned covariance structure of the entire population. This allows the system to rapidly narrow down the user's likely position in the latent space with minimal upfront cognitive burden.
Exploiting Structural and Semantic Hierarchies
When explicit interaction identifiers (IDs) are missing, cold-start models can also exploit the structural relationships between entities. An item does not exist in a vacuum; it belongs to a brand, a category, or a specific merchant.
Kui Zhao et al. explored this structural inheritance in Learning and Transferring IDs Representation in E-commerce. By embedding users, items, brands, and stores into a single, shared low-dimensional semantic space, a new item can leverage the historical behavior tied to its parent brand or category. Even if the item ID itself has zero interactions, its structural neighbors pull it into a highly relevant neighborhood within the latent space.
On the purely perceptual side, systems can completely replace discrete IDs with continuous semantic representations. In Telepath: Understanding Users from a Human Vision Perspective in Large-Scale Recommender Systems, Yu Wang et al. bypassed traditional item identifiers by using deep learning to simulate human vision. Because the model maps raw visual appearance directly to user interest vectors, a new item is never truly "cold"—its visual features instantly anchor it to the preferences of users who have interacted with visually similar items in the past.