Most recommendation systems today rely on a simple mathematical operation: the dot product. By calculating the inner product between a user vector and an item vector, models like Matrix Factorization estimate preference. However, as Lina Yao and Chanyoung Park have observed, this approach has a fundamental geometric flaw. Dot products do not satisfy the triangle inequality, a mathematical property that is essential for a latent space to be logically consistent.
When a model violates the triangle inequality, it loses the ability to perform "collaborative transitivity." If a user likes Item A and Item B, a consistent geometric space should imply that Item A and Item B are related. Without metric properties, a model can place Item A and Item B on opposite sides of the universe while still keeping them both "close" to the user via the dot product. This inconsistency prevents the system from generalizing to unobserved interactions, especially in sparse datasets.
The Geometry of Consistency
To fix this, researchers are moving toward Metric Learning. In these models, the "score" is not a dot product but a distance (usually Euclidean) in a latent space. Because distances must satisfy the triangle inequality ($d(x, z) \leq d(x, y) + d(y, z)$), the model is forced to organize items into coherent clusters.
As Ruining He and Julian McAuley demonstrate with TransRec, this metricity allows for powerful relational inference. If we know a user moves from Item A to Item B, and another user moves from Item B to Item C, a metric space naturally suggests a transition from A to C. This "transitivity" acts as a bridge across sparse data, allowing the model to infer relationships between items that have never appeared together in the same history.
Overcoming the Rigidity of Points
While simple distance metrics improve consistency, representing a user as a single point in space can be too "rigid." A user who likes both horror movies and romantic comedies might be forced into a single point halfway between the two, which represents neither.
To solve this, Chen Ma and Ruiming Tang introduced PMLAM, which replaces fixed points with probabilistic distributions (Gaussian clouds). Instead of measuring the distance between two dots, the model measures the Wasserstein distance between two distributions. This allows the model to capture uncertainty and "diverse preferences"—the width of the Gaussian represents how broad or specific a user's taste is.
Similarly, Zhiwei Liu and Ziwei Fan utilized this stochastic approach in STOSA to handle the inherent noise in sequential behavior. By treating items as distributions, the model can maintain the triangle inequality while accounting for the fact that a "click" doesn't always represent a perfect match for a user's long-term intent.
Solving the Cold Start Problem
The primary benefit of this shift from dot products to metrics is robustness. Traditional similarity measures, like those analyzed by Guibing Guo, often fail when users have only a few ratings, leading to "chance correlations."
By contrast, metric learning frameworks like METAS by Dongha Lee and Chanyoung Park use the geometric structure of the space to "pull" new items or users into the correct clusters based on their shared relationships. As Yuanshuai Luo argues, this structured embedding space is significantly more effective at addressing the cold start problem. When the latent space is a true metric space, every new piece of data informs the entire structure through the transitive property, making the system far more than the sum of its individual interactions.