src/methods/pca_dir.py

"""M3 — PCADir: PC1 of (H_pos - H_neg) activation matrix."""

import numpy as np
from sklearn.decomposition import PCA

from src.methods.base import SteeringMethod


class PCADir(SteeringMethod):
    """PCADir — First principal component of contrastive activations."""

    @property
    def name(self) -> str:
        return "PCADir"

    @property
    def method_id(self) -> str:
        return "M3"

    def extract_vector(
        self,
        h_pos: np.ndarray,
        h_neg: np.ndarray,
        **kwargs,
    ) -> np.ndarray:
        """Compute PC1 of the concatenated [H_pos; H_neg] activation matrix.

        This captures the direction of maximum variance, which corresponds to
        the axis separating the positive and negative distributions.

        Args:
            h_pos: (N, d) positive activations
            h_neg: (N, d) negative activations

        Returns:
            (d,) first principal component direction (unit norm)
        """
        # Concatenate and find PC1 of the combined activations
        H = np.concatenate([h_pos, h_neg], axis=0)  # (2N, d)
        pca = PCA(n_components=1)
        pca.fit(H)
        v = pca.components_[0]  # (d,)

        # Orient: ensure positive dot product with mean diff
        mean_diff = h_pos.mean(axis=0) - h_neg.mean(axis=0)
        if np.dot(v, mean_diff) < 0:
            v = -v

        return v