Source code for simba.analog_discovery.simba_analog_discovery

import numpy as np




class AnalogDiscovery:


    @staticmethod
    def compute_ranking(similarities_mces, similarities_ed, max_value_2_int=5):
        """
        based on mces and edit distance rerank.
        If 2 matches have the same mces, choose the one with lowest edit distance
        """
        similarities_mces_integer = np.round(similarities_mces)
        # Preallocate the ranking array with the same shape as similarities1.
        ranking_total = np.zeros(similarities_mces.shape, dtype=int)

        # Process each row (or each set of values) individually.
        for row_index, (row_sim, row_int, row_int2) in enumerate(
            zip(
                similarities_mces,
                similarities_mces_integer,
                similarities_ed,
                strict=False,
            )
        ):
            # Use lexsort with a composite key:
            #   - Primary: similarities1_integer (ascending)
            #   - Secondary: similarities2_integer (ascending)
            #   - Tertiary: similarities1 (descending, so use -row_sim)
            #
            # Note: np.lexsort uses the last key as the primary key.
            sorted_indices = np.lexsort((row_sim, row_int2, row_int))

            # Now assign ranking values based on sorted order.
            # Here the best (first in sorted_indices) gets rank 0,
            # the next gets rank 1, etc.
            ranking = np.empty_like(sorted_indices)
            ranking[sorted_indices] = np.arange(len(row_sim))

            # Store the ranking for this row.
            ranking_total[row_index] = ranking

        # normalizing
        ranking_total = 1 - ranking_total / ranking_total.shape[1]
        return ranking_total