siamese_from_correlation_matrix

What problem does it solve?

Derives training signals for siamese or metric-learning models directly from an embedding correlation matrix, eliminating the need for external labels.

Core Features & Use Cases

• Correlation-to-training target generation: Converts an N×N correlation matrix (e.g., cosine/dot-product structure) into regression-style targets for pairwise similarity learning.
• Structure-aware hybrid supervision: Augments correlation regression with cluster membership signals (e.g., GMM/HDBSCAN) and decorrelation-based boundary/ordinal guidance.
• Model refinement and interpretability: Produces a refined embedding space and supports integrated visualization (sorted correlation heatmaps, embedding clouds, and distribution plots) to verify learned structure.

Quick Start

Use siamese_from_correlation_matrix to train a contrastive/metric-learning model from a pre-computed embedding correlation matrix you already have, and generate a refined embedding model that preserves correlation and cluster structure.