Merge branch 'knn-graph-clustering' into 'master'

knngraph process constructs a geometric graph based on nearest neighbours

See merge request postprocessor!11

core/processes/knngraph.py

+#!/usr/bin/env python3
+
+import pandas as pd
+from sklearn.metrics.pairwise import euclidean_distances
+import igraph as ig
+
+from agora.abc import ParametersABC
+from postprocessor.core.abc import PostProcessABC
+
+
+class knngraphParameters(ParametersABC):
+    """
+    Parameters for the 'knngraph' process.
+    """
+
+    _defaults = {
+        "n_neighbors": 10,
+    }
+
+
+class knngraph(PostProcessABC):
+    """
+    Process to get geometric graph from data/features, based on k-nearest
+    neighbours.
+
+    Methods
+    -------
+    run(signal: pd.DataFrame)
+        Get graph.
+    """
+
+    def __init__(self, parameters: knngraphParameters):
+        super().__init__(parameters)
+
+    def run(self, signal: pd.DataFrame):
+        """Get graph.
+
+        Parameters
+        ----------
+        signal : pd.DataFrame
+            Feature matrix.
+        """
+        distance_matrix = euclidean_distances(signal)
+        distance_matrix_pruned = graph_prune(distance_matrix, self.n_neighbours)
+        graph = ig.Graph.Weighted_Adjacency(
+            distance_matrix_pruned.tolist(), mode="undirected"
+        )
+        graph.vs["strain"] = signal.index.get_level_values("strain")
+        return graph
+
+
+def graph_prune(distanceMatrix, neighbours):
+    """
+    Prunes a complete graph (input distance matrix), keeping at least a
+    specified number of neighbours for each node.
+
+    Parameters:
+    -----------
+    distanceMatrix = 2D numpy array
+    neighbours = integer
+
+    Return: Dij_pruned, a 2D numpy array, represents distance matrix of pruned
+            graph
+    """
+    Dij_temp = distanceMatrix
+    Adj = np.zeros(distanceMatrix.shape)
+    for ii in range(distanceMatrix.shape[0]):
+        idx = np.argsort(Dij_temp[ii, :])
+        Adj[ii, idx[1]] = 1
+        Adj[idx[1], ii] = 1
+        for jj in range(neighbours):
+            Adj[ii, idx[jj]] = 1
+            Adj[idx[jj], ii] = 1
+    Dij_pruned = Dij_temp * Adj
+    return Dij_pruned