median-bifurcation

What problem does it solve?

Median bifurcation creates discriminative negative signal by generating both sides of a distinction instead of training only on positive examples, improving learning and downstream decision boundaries.

Core Features & Use Cases

• Balanced median partitioning: Splits any continuous axis at a robust median (or equivalent), producing equal-sized partitions and avoiding class-imbalance drift.
• Hard negatives baked in: Produces both the “entailed/positive” and “non-entailed/negative” halves explicitly at data-prep time rather than mining negatives later.
• Recursive cell refinement: Applies the split repeatedly across orthogonal axes to grow epistemic granularity (2^n cells) with O(n axes) labeling cost.
• Works across domains: Supports training-data preparation (SPO-style entailment), feature engineering/clustering, dynamic batching, and ANOVA-style factor analysis.
• Inference-time pruning: Drops the unwanted partition at runtime so only the relevant side is emitted.

Quick Start

Use median-bifurcation to split your samples into above/below-median (or entailed/non-entailed) groups, include both halves during training, and emit only the relevant half at inference.