Siamese Network

A Siamese network consists of two identical subnetworks that process two distinct inputs in parallel. Each branch computes an embedding for its respective input. These embeddings are then fed into a distance metric layer—such as Euclidean distance or cosine similarity—to produce a final similarity score. This design is fundamentally different from a single-stream classifier because it learns a relative relationship between inputs rather than absolute class boundaries.

The core constraint of a Siamese architecture is that both branches share the exact same weights and are updated simultaneously during backpropagation. This weight-tying guarantees that the network processes both inputs identically, ensuring that the resulting embeddings reside in the same latent space. Without this symmetry, the network could learn arbitrary mappings for each input, making distance comparisons meaningless. The shared-weight constraint is what enforces the learning of a true similarity metric.

Siamese networks are typically trained using a contrastive loss or triplet loss, not standard cross-entropy. Contrastive loss operates on pairs of inputs and minimizes the distance between embeddings of similar pairs while maximizing the distance for dissimilar pairs beyond a defined margin. The formula is:

• L = (1 - Y) * ½(D_w)² + Y * ½{max(0, m - D_w)}² where Y is the binary label (0 for similar, 1 for dissimilar), D_w is the distance, and m is the margin. This loss directly optimizes the embedding space for nearest-neighbor retrieval.

Because a Siamese network learns a similarity function rather than memorizing class labels, it can compare a new user or item against known examples without any retraining. For a cold-start user with only demographic data, the network computes an embedding and measures its distance to item embeddings in the catalog. The items with the smallest distance are recommended immediately. This zero-shot matching is the primary reason Siamese architectures are deployed for cold-start problem mitigation.

Siamese networks excel at one-shot learning, where a model must generalize from a single example of a new class. In a retail context, a new product with only one interaction can be embedded and compared against all user embeddings to find the most similar audience segment. This capability stems from the network's training regime, which is structured as a series of pairwise verification tasks rather than multi-class classification, forcing it to learn transferable comparison heuristics.

The output of each branch is a dense, low-dimensional vector—an embedding—that captures the semantic essence of the input. After training, the embedding space exhibits a critical property: semantically similar inputs cluster together, while dissimilar inputs are pushed apart. This structured latent space enables efficient approximate nearest neighbor (ANN) search using libraries like FAISS or Annoy, allowing real-time retrieval of the top-K most similar items for any cold-start query vector.

The mathematical objective that drives a Siamese Network's training. It minimizes the distance between similar pairs (positive samples) and maximizes the distance between dissimilar pairs (negative samples) up to a defined margin. This loss function directly shapes the embedding space to reflect semantic similarity, ensuring that a new user's sparse profile is pulled close to matching item attributes while being pushed away from irrelevant ones.

An extension of the Siamese architecture that processes three inputs simultaneously: an anchor, a positive example, and a negative example. The triplet loss ensures the distance between the anchor and positive is smaller than the distance between the anchor and negative by a specific margin. This is particularly effective for fine-grained ranking in cold-start scenarios, teaching the model to discern subtle preference differences from minimal initial data.

The ability to recognize or classify a new entity after seeing only a single example. Siamese Networks are foundational to this capability because they learn a similarity function rather than memorizing classes. In a retail cold start, this allows the system to match a new user to a product category after just one click, comparing the interaction embedding directly to item embeddings without requiring a retrained classifier.

The final layer of a Siamese Network computes a distance score between two output vectors. Euclidean distance measures the straight-line magnitude in embedding space, while cosine similarity measures the angle between vectors, ignoring magnitude. Cosine similarity is often preferred for cold-start matching because it focuses on directional preference alignment rather than intensity, making it robust to the sparse, noisy signals from new user sessions.

The defining structural constraint of a Siamese Network: both input branches are processed by identical subnetworks with tied weights. This guarantees that similar inputs are mapped to similar regions of the embedding space using the same transformation function. Without weight sharing, the network could learn arbitrary, inconsistent mappings that fail to generalize to the zero-shot comparison required for cold-start item matching.

Siamese Networks were first introduced in the 1990s by Bromley et al. for offline signature verification, comparing a test signature against a known reference to detect forgery. This historical application mirrors the modern cold-start use case: determining if a new, unseen input (a new user's behavior) is genuinely similar to a known reference (an item profile), making it a natural fit for zero-shot identity and preference matching.