okstra 0.108.0 → 0.109.0

package/runtime/python/okstra_vendor/networkx/algorithms/community/lukes.py

@@ -0,0 +1,227 @@
+"""Lukes Algorithm for exact optimal weighted tree partitioning."""
+
+from copy import deepcopy
+from functools import lru_cache
+from random import choice
+
+import networkx as nx
+from networkx.utils import not_implemented_for
+
+__all__ = ["lukes_partitioning"]
+
+D_EDGE_W = "weight"
+D_EDGE_VALUE = 1.0
+D_NODE_W = "weight"
+D_NODE_VALUE = 1
+PKEY = "partitions"
+CLUSTER_EVAL_CACHE_SIZE = 2048
+
+
+def _split_n_from(n, min_size_of_first_part):
+    # splits j in two parts of which the first is at least
+    # the second argument
+    assert n >= min_size_of_first_part
+    for p1 in range(min_size_of_first_part, n + 1):
+        yield p1, n - p1
+
+
+@nx._dispatchable(node_attrs="node_weight", edge_attrs="edge_weight")
+def lukes_partitioning(G, max_size, node_weight=None, edge_weight=None):
+    """Optimal partitioning of a weighted tree using the Lukes algorithm.
+
+    This algorithm partitions a connected, acyclic graph featuring integer
+    node weights and float edge weights. The resulting clusters are such
+    that the total weight of the nodes in each cluster does not exceed
+    max_size and that the weight of the edges that are cut by the partition
+    is minimum. The algorithm is based on [1]_.
+
+    Parameters
+    ----------
+    G : NetworkX graph
+
+    max_size : int
+        Maximum weight a partition can have in terms of sum of
+        node_weight for all nodes in the partition
+
+    edge_weight : key
+        Edge data key to use as weight. If None, the weights are all
+        set to one.
+
+    node_weight : key
+        Node data key to use as weight. If None, the weights are all
+        set to one. The data must be int.
+
+    Returns
+    -------
+    partition : list
+        A list of sets of nodes representing the clusters of the
+        partition.
+
+    Raises
+    ------
+    NotATree
+        If G is not a tree.
+    TypeError
+        If any of the values of node_weight is not int.
+
+    References
+    ----------
+    .. [1] Lukes, J. A. (1974).
+       "Efficient Algorithm for the Partitioning of Trees."
+       IBM Journal of Research and Development, 18(3), 217–224.
+
+    """
+    # First sanity check and tree preparation
+    if not nx.is_tree(G):
+        raise nx.NotATree("lukes_partitioning works only on trees")
+    else:
+        if nx.is_directed(G):
+            root = [n for n, d in G.in_degree() if d == 0]
+            assert len(root) == 1
+            root = root[0]
+            t_G = deepcopy(G)
+        else:
+            root = choice(list(G.nodes))
+            # this has the desirable side effect of not inheriting attributes
+            t_G = nx.dfs_tree(G, root)
+
+    # Since we do not want to screw up the original graph,
+    # if we have a blank attribute, we make a deepcopy
+    if edge_weight is None or node_weight is None:
+        safe_G = deepcopy(G)
+        if edge_weight is None:
+            nx.set_edge_attributes(safe_G, D_EDGE_VALUE, D_EDGE_W)
+            edge_weight = D_EDGE_W
+        if node_weight is None:
+            nx.set_node_attributes(safe_G, D_NODE_VALUE, D_NODE_W)
+            node_weight = D_NODE_W
+    else:
+        safe_G = G
+
+    # Second sanity check
+    # The values of node_weight MUST BE int.
+    # I cannot see any room for duck typing without incurring serious
+    # danger of subtle bugs.
+    all_n_attr = nx.get_node_attributes(safe_G, node_weight).values()
+    for x in all_n_attr:
+        if not isinstance(x, int):
+            raise TypeError(
+                "lukes_partitioning needs integer "
+                f"values for node_weight ({node_weight})"
+            )
+
+    # SUBROUTINES -----------------------
+    # these functions are defined here for two reasons:
+    # - brevity: we can leverage global "safe_G"
+    # - caching: signatures are hashable
+
+    @not_implemented_for("undirected")
+    # this is intended to be called only on t_G
+    def _leaves(gr):
+        for x in gr.nodes:
+            if not nx.descendants(gr, x):
+                yield x
+
+    @not_implemented_for("undirected")
+    def _a_parent_of_leaves_only(gr):
+        tleaves = set(_leaves(gr))
+        for n in set(gr.nodes) - tleaves:
+            if all(x in tleaves for x in nx.descendants(gr, n)):
+                return n
+
+    @lru_cache(CLUSTER_EVAL_CACHE_SIZE)
+    def _value_of_cluster(cluster):
+        valid_edges = [e for e in safe_G.edges if e[0] in cluster and e[1] in cluster]
+        return sum(safe_G.edges[e][edge_weight] for e in valid_edges)
+
+    def _value_of_partition(partition):
+        return sum(_value_of_cluster(frozenset(c)) for c in partition)
+
+    @lru_cache(CLUSTER_EVAL_CACHE_SIZE)
+    def _weight_of_cluster(cluster):
+        return sum(safe_G.nodes[n][node_weight] for n in cluster)
+
+    def _pivot(partition, node):
+        ccx = [c for c in partition if node in c]
+        assert len(ccx) == 1
+        return ccx[0]
+
+    def _concatenate_or_merge(partition_1, partition_2, x, i, ref_weight):
+        ccx = _pivot(partition_1, x)
+        cci = _pivot(partition_2, i)
+        merged_xi = ccx.union(cci)
+
+        # We first check if we can do the merge.
+        # If so, we do the actual calculations, otherwise we concatenate
+        if _weight_of_cluster(frozenset(merged_xi)) <= ref_weight:
+            cp1 = list(filter(lambda x: x != ccx, partition_1))
+            cp2 = list(filter(lambda x: x != cci, partition_2))
+
+            option_2 = [merged_xi] + cp1 + cp2
+            return option_2, _value_of_partition(option_2)
+        else:
+            option_1 = partition_1 + partition_2
+            return option_1, _value_of_partition(option_1)
+
+    # INITIALIZATION -----------------------
+    leaves = set(_leaves(t_G))
+    for lv in leaves:
+        t_G.nodes[lv][PKEY] = {}
+        slot = safe_G.nodes[lv][node_weight]
+        t_G.nodes[lv][PKEY][slot] = [{lv}]
+        t_G.nodes[lv][PKEY][0] = [{lv}]
+
+    for inner in [x for x in t_G.nodes if x not in leaves]:
+        t_G.nodes[inner][PKEY] = {}
+        slot = safe_G.nodes[inner][node_weight]
+        t_G.nodes[inner][PKEY][slot] = [{inner}]
+    nx._clear_cache(t_G)
+
+    # CORE ALGORITHM -----------------------
+    while True:
+        x_node = _a_parent_of_leaves_only(t_G)
+        weight_of_x = safe_G.nodes[x_node][node_weight]
+        best_value = 0
+        best_partition = None
+        bp_buffer = {}
+        x_descendants = nx.descendants(t_G, x_node)
+        for i_node in x_descendants:
+            for j in range(weight_of_x, max_size + 1):
+                for a, b in _split_n_from(j, weight_of_x):
+                    if (
+                        a not in t_G.nodes[x_node][PKEY]
+                        or b not in t_G.nodes[i_node][PKEY]
+                    ):
+                        # it's not possible to form this particular weight sum
+                        continue
+
+                    part1 = t_G.nodes[x_node][PKEY][a]
+                    part2 = t_G.nodes[i_node][PKEY][b]
+                    part, value = _concatenate_or_merge(part1, part2, x_node, i_node, j)
+
+                    if j not in bp_buffer or bp_buffer[j][1] < value:
+                        # we annotate in the buffer the best partition for j
+                        bp_buffer[j] = part, value
+
+                    # we also keep track of the overall best partition
+                    if best_value <= value:
+                        best_value = value
+                        best_partition = part
+
+            # as illustrated in Lukes, once we finished a child, we can
+            # discharge the partitions we found into the graph
+            # (the key phrase is make all x == x')
+            # so that they are used by the subsequent children
+            for w, (best_part_for_vl, vl) in bp_buffer.items():
+                t_G.nodes[x_node][PKEY][w] = best_part_for_vl
+            bp_buffer.clear()
+
+        # the absolute best partition for this node
+        # across all weights has to be stored at 0
+        t_G.nodes[x_node][PKEY][0] = best_partition
+        t_G.remove_nodes_from(x_descendants)
+
+        if x_node == root:
+            # the 0-labeled partition of root
+            # is the optimal one for the whole tree
+            return t_G.nodes[root][PKEY][0]

package/runtime/python/okstra_vendor/networkx/algorithms/community/modularity_max.py

@@ -0,0 +1,452 @@
+"""Functions for detecting communities based on modularity."""
+
+import random
+from collections import defaultdict
+from copy import deepcopy
+
+import networkx as nx
+from networkx.algorithms.community.quality import modularity
+from networkx.utils.mapped_queue import MappedQueue
+
+__all__ = [
+    "greedy_modularity_communities",
+    "naive_greedy_modularity_communities",
+]
+
+
+def _greedy_modularity_communities_generator(G, weight=None, resolution=1):
+    r"""Yield community partitions of G and the modularity change at each step.
+
+    This function performs Clauset-Newman-Moore greedy modularity maximization [2]_
+    At each step of the process it yields the change in modularity that will occur in
+    the next step followed by yielding the new community partition after that step.
+
+    Greedy modularity maximization begins with each node in its own community
+    and repeatedly joins the pair of communities that lead to the largest
+    modularity until one community contains all nodes (the partition has one set).
+
+    This function maximizes the generalized modularity, where `resolution`
+    is the resolution parameter, often expressed as $\gamma$.
+    See :func:`~networkx.algorithms.community.quality.modularity`.
+
+    Parameters
+    ----------
+    G : NetworkX graph
+
+    weight : string or None, optional (default=None)
+        The name of an edge attribute that holds the numerical value used
+        as a weight.  If None, then each edge has weight 1.
+        The degree is the sum of the edge weights adjacent to the node.
+
+    resolution : float (default=1)
+        If resolution is less than 1, modularity favors larger communities.
+        Greater than 1 favors smaller communities.
+
+    Yields
+    ------
+    Alternating yield statements produce the following two objects:
+
+    communities: dict_values
+        A dict_values of frozensets of nodes, one for each community.
+        This represents a partition of the nodes of the graph into communities.
+        The first yield is the partition with each node in its own community.
+
+    dq: float
+        The change in modularity when merging the next two communities
+        that leads to the largest modularity.
+
+    See Also
+    --------
+    modularity
+
+    References
+    ----------
+    .. [1] Newman, M. E. J. "Networks: An Introduction", page 224
+       Oxford University Press 2011.
+    .. [2] Clauset, A., Newman, M. E., & Moore, C.
+       "Finding community structure in very large networks."
+       Physical Review E 70(6), 2004.
+    .. [3] Reichardt and Bornholdt "Statistical Mechanics of Community
+       Detection" Phys. Rev. E74, 2006.
+    .. [4] Newman, M. E. J."Analysis of weighted networks"
+       Physical Review E 70(5 Pt 2):056131, 2004.
+    """
+    directed = G.is_directed()
+    N = G.number_of_nodes()
+
+    # Count edges (or the sum of edge-weights for weighted graphs)
+    m = G.size(weight)
+    q0 = 1 / m
+
+    # Calculate degrees (notation from the papers)
+    # a : the fraction of (weighted) out-degree for each node
+    # b : the fraction of (weighted) in-degree for each node
+    if directed:
+        a = {node: deg_out * q0 for node, deg_out in G.out_degree(weight=weight)}
+        b = {node: deg_in * q0 for node, deg_in in G.in_degree(weight=weight)}
+    else:
+        a = b = {node: deg * q0 * 0.5 for node, deg in G.degree(weight=weight)}
+
+    # this preliminary step collects the edge weights for each node pair
+    # It handles multigraph and digraph and works fine for graph.
+    dq_dict = defaultdict(lambda: defaultdict(float))
+    for u, v, wt in G.edges(data=weight, default=1):
+        if u == v:
+            continue
+        dq_dict[u][v] += wt
+        dq_dict[v][u] += wt
+
+    # now scale and subtract the expected edge-weights term
+    for u, nbrdict in dq_dict.items():
+        for v, wt in nbrdict.items():
+            dq_dict[u][v] = q0 * wt - resolution * (a[u] * b[v] + b[u] * a[v])
+
+    # Use -dq to get a max_heap instead of a min_heap
+    # dq_heap holds a heap for each node's neighbors
+    dq_heap = {u: MappedQueue({(u, v): -dq for v, dq in dq_dict[u].items()}) for u in G}
+    # H -> all_dq_heap holds a heap with the best items for each node
+    H = MappedQueue([dq_heap[n].heap[0] for n in G if len(dq_heap[n]) > 0])
+
+    # Initialize single-node communities
+    communities = {n: frozenset([n]) for n in G}
+    yield communities.values()
+
+    # Merge the two communities that lead to the largest modularity
+    while len(H) > 1:
+        # Find best merge
+        # Remove from heap of row maxes
+        # Ties will be broken by choosing the pair with lowest min community id
+        try:
+            negdq, u, v = H.pop()
+        except IndexError:
+            break
+        dq = -negdq
+        yield dq
+        # Remove best merge from row u heap
+        dq_heap[u].pop()
+        # Push new row max onto H
+        if len(dq_heap[u]) > 0:
+            H.push(dq_heap[u].heap[0])
+        # If this element was also at the root of row v, we need to remove the
+        # duplicate entry from H
+        if dq_heap[v].heap[0] == (v, u):
+            H.remove((v, u))
+            # Remove best merge from row v heap
+            dq_heap[v].remove((v, u))
+            # Push new row max onto H
+            if len(dq_heap[v]) > 0:
+                H.push(dq_heap[v].heap[0])
+        else:
+            # Duplicate wasn't in H, just remove from row v heap
+            dq_heap[v].remove((v, u))
+
+        # Perform merge
+        communities[v] = frozenset(communities[u] | communities[v])
+        del communities[u]
+
+        # Get neighbor communities connected to the merged communities
+        u_nbrs = set(dq_dict[u])
+        v_nbrs = set(dq_dict[v])
+        all_nbrs = (u_nbrs | v_nbrs) - {u, v}
+        both_nbrs = u_nbrs & v_nbrs
+        # Update dq for merge of u into v
+        for w in all_nbrs:
+            # Calculate new dq value
+            if w in both_nbrs:
+                dq_vw = dq_dict[v][w] + dq_dict[u][w]
+            elif w in v_nbrs:
+                dq_vw = dq_dict[v][w] - resolution * (a[u] * b[w] + a[w] * b[u])
+            else:  # w in u_nbrs
+                dq_vw = dq_dict[u][w] - resolution * (a[v] * b[w] + a[w] * b[v])
+            # Update rows v and w
+            for row, col in [(v, w), (w, v)]:
+                dq_heap_row = dq_heap[row]
+                # Update dict for v,w only (u is removed below)
+                dq_dict[row][col] = dq_vw
+                # Save old max of per-row heap
+                if len(dq_heap_row) > 0:
+                    d_oldmax = dq_heap_row.heap[0]
+                else:
+                    d_oldmax = None
+                # Add/update heaps
+                d = (row, col)
+                d_negdq = -dq_vw
+                # Save old value for finding heap index
+                if w in v_nbrs:
+                    # Update existing element in per-row heap
+                    dq_heap_row.update(d, d, priority=d_negdq)
+                else:
+                    # We're creating a new nonzero element, add to heap
+                    dq_heap_row.push(d, priority=d_negdq)
+                # Update heap of row maxes if necessary
+                if d_oldmax is None:
+                    # No entries previously in this row, push new max
+                    H.push(d, priority=d_negdq)
+                else:
+                    # We've updated an entry in this row, has the max changed?
+                    row_max = dq_heap_row.heap[0]
+                    if d_oldmax != row_max or d_oldmax.priority != row_max.priority:
+                        H.update(d_oldmax, row_max)
+
+        # Remove row/col u from dq_dict matrix
+        for w in dq_dict[u]:
+            # Remove from dict
+            dq_old = dq_dict[w][u]
+            del dq_dict[w][u]
+            # Remove from heaps if we haven't already
+            if w != v:
+                # Remove both row and column
+                for row, col in [(w, u), (u, w)]:
+                    dq_heap_row = dq_heap[row]
+                    # Check if replaced dq is row max
+                    d_old = (row, col)
+                    if dq_heap_row.heap[0] == d_old:
+                        # Update per-row heap and heap of row maxes
+                        dq_heap_row.remove(d_old)
+                        H.remove(d_old)
+                        # Update row max
+                        if len(dq_heap_row) > 0:
+                            H.push(dq_heap_row.heap[0])
+                    else:
+                        # Only update per-row heap
+                        dq_heap_row.remove(d_old)
+
+        del dq_dict[u]
+        # Mark row u as deleted, but keep placeholder
+        dq_heap[u] = MappedQueue()
+        # Merge u into v and update a
+        a[v] += a[u]
+        a[u] = 0
+        if directed:
+            b[v] += b[u]
+            b[u] = 0
+
+        yield communities.values()
+
+
+@nx._dispatchable(edge_attrs="weight")
+def greedy_modularity_communities(
+    G,
+    weight=None,
+    resolution=1,
+    cutoff=1,
+    best_n=None,
+):
+    r"""Find communities in G using greedy modularity maximization.
+
+    This function uses Clauset-Newman-Moore greedy modularity maximization [2]_
+    to find the community partition with the largest modularity.
+
+    Greedy modularity maximization begins with each node in its own community
+    and repeatedly joins the pair of communities that lead to the largest
+    modularity until no further increase in modularity is possible (a maximum).
+    Two keyword arguments adjust the stopping condition. `cutoff` is a lower
+    limit on the number of communities so you can stop the process before
+    reaching a maximum (used to save computation time). `best_n` is an upper
+    limit on the number of communities so you can make the process continue
+    until at most n communities remain even if the maximum modularity occurs
+    for more. To obtain exactly n communities, set both `cutoff` and `best_n` to n.
+
+    This function maximizes the generalized modularity, where `resolution`
+    is the resolution parameter, often expressed as $\gamma$.
+    See :func:`~networkx.algorithms.community.quality.modularity`.
+
+    Parameters
+    ----------
+    G : NetworkX graph
+
+    weight : string or None, optional (default=None)
+        The name of an edge attribute that holds the numerical value used
+        as a weight.  If None, then each edge has weight 1.
+        The degree is the sum of the edge weights adjacent to the node.
+
+    resolution : float, optional (default=1)
+        If resolution is less than 1, modularity favors larger communities.
+        Greater than 1 favors smaller communities.
+
+    cutoff : int, optional (default=1)
+        A minimum number of communities below which the merging process stops.
+        The process stops at this number of communities even if modularity
+        is not maximized. The goal is to let the user stop the process early.
+        The process stops before the cutoff if it finds a maximum of modularity.
+
+    best_n : int or None, optional (default=None)
+        A maximum number of communities above which the merging process will
+        not stop. This forces community merging to continue after modularity
+        starts to decrease until `best_n` communities remain.
+        If ``None``, don't force it to continue beyond a maximum.
+
+    Raises
+    ------
+    ValueError : If the `cutoff` or `best_n`  value is not in the range
+        ``[1, G.number_of_nodes()]``, or if `best_n` < `cutoff`.
+
+    Returns
+    -------
+    communities: list
+        A list of frozensets of nodes, one for each community.
+        Sorted by length with largest communities first.
+
+    Examples
+    --------
+    >>> G = nx.karate_club_graph()
+    >>> c = nx.community.greedy_modularity_communities(G)
+    >>> sorted(c[0])
+    [8, 14, 15, 18, 20, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33]
+
+    See Also
+    --------
+    modularity
+
+    References
+    ----------
+    .. [1] Newman, M. E. J. "Networks: An Introduction", page 224
+       Oxford University Press 2011.
+    .. [2] Clauset, A., Newman, M. E., & Moore, C.
+       "Finding community structure in very large networks."
+       Physical Review E 70(6), 2004.
+    .. [3] Reichardt and Bornholdt "Statistical Mechanics of Community
+       Detection" Phys. Rev. E74, 2006.
+    .. [4] Newman, M. E. J."Analysis of weighted networks"
+       Physical Review E 70(5 Pt 2):056131, 2004.
+    """
+    if not G.size():
+        return [{n} for n in G]
+
+    if (cutoff < 1) or (cutoff > G.number_of_nodes()):
+        raise ValueError(f"cutoff must be between 1 and {len(G)}. Got {cutoff}.")
+    if best_n is not None:
+        if (best_n < 1) or (best_n > G.number_of_nodes()):
+            raise ValueError(f"best_n must be between 1 and {len(G)}. Got {best_n}.")
+        if best_n < cutoff:
+            raise ValueError(f"Must have best_n >= cutoff. Got {best_n} < {cutoff}")
+        if best_n == 1:
+            return [set(G)]
+    else:
+        best_n = G.number_of_nodes()
+
+    # retrieve generator object to construct output
+    community_gen = _greedy_modularity_communities_generator(
+        G, weight=weight, resolution=resolution
+    )
+
+    # construct the first best community
+    communities = next(community_gen)
+
+    # continue merging communities until one of the breaking criteria is satisfied
+    while len(communities) > cutoff:
+        try:
+            dq = next(community_gen)
+        # StopIteration occurs when communities are the connected components
+        except StopIteration:
+            communities = sorted(communities, key=len, reverse=True)
+            # if best_n requires more merging, merge big sets for highest modularity
+            while len(communities) > best_n:
+                comm1, comm2, *rest = communities
+                communities = [comm1 ^ comm2]
+                communities.extend(rest)
+            return communities
+
+        # keep going unless max_mod is reached or best_n says to merge more
+        if dq < 0 and len(communities) <= best_n:
+            break
+        communities = next(community_gen)
+
+    return sorted(communities, key=len, reverse=True)
+
+
+@nx.utils.not_implemented_for("directed")
+@nx.utils.not_implemented_for("multigraph")
+@nx._dispatchable(edge_attrs="weight")
+def naive_greedy_modularity_communities(G, resolution=1, weight=None):
+    r"""Find communities in G using greedy modularity maximization.
+
+    This implementation is O(n^4), much slower than alternatives, but it is
+    provided as an easy-to-understand reference implementation.
+
+    Greedy modularity maximization begins with each node in its own community
+    and joins the pair of communities that most increases modularity until no
+    such pair exists.
+
+    This function maximizes the generalized modularity, where `resolution`
+    is the resolution parameter, often expressed as $\gamma$.
+    See :func:`~networkx.algorithms.community.quality.modularity`.
+
+    Parameters
+    ----------
+    G : NetworkX graph
+        Graph must be simple and undirected.
+
+    resolution : float (default=1)
+        If resolution is less than 1, modularity favors larger communities.
+        Greater than 1 favors smaller communities.
+
+    weight : string or None, optional (default=None)
+        The name of an edge attribute that holds the numerical value used
+        as a weight.  If None, then each edge has weight 1.
+        The degree is the sum of the edge weights adjacent to the node.
+
+    Returns
+    -------
+    list
+        A list of sets of nodes, one for each community.
+        Sorted by length with largest communities first.
+
+    Examples
+    --------
+    >>> G = nx.karate_club_graph()
+    >>> c = nx.community.naive_greedy_modularity_communities(G)
+    >>> sorted(c[0])
+    [8, 14, 15, 18, 20, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33]
+
+    See Also
+    --------
+    greedy_modularity_communities
+    modularity
+    """
+    # First create one community for each node
+    communities = [frozenset([u]) for u in G.nodes()]
+    # Track merges
+    merges = []
+    # Greedily merge communities until no improvement is possible
+    old_modularity = None
+    new_modularity = modularity(G, communities, resolution=resolution, weight=weight)
+    while old_modularity is None or new_modularity > old_modularity:
+        # Save modularity for comparison
+        old_modularity = new_modularity
+        # Find best pair to merge
+        trial_communities = list(communities)
+        to_merge = None
+        for i, u in enumerate(communities):
+            for j, v in enumerate(communities):
+                # Skip i==j and empty communities
+                if j <= i or len(u) == 0 or len(v) == 0:
+                    continue
+                # Merge communities u and v
+                trial_communities[j] = u | v
+                trial_communities[i] = frozenset([])
+                trial_modularity = modularity(
+                    G, trial_communities, resolution=resolution, weight=weight
+                )
+                if trial_modularity >= new_modularity:
+                    # Check if strictly better or tie
+                    if trial_modularity > new_modularity:
+                        # Found new best, save modularity and group indexes
+                        new_modularity = trial_modularity
+                        to_merge = (i, j, new_modularity - old_modularity)
+                    elif to_merge and min(i, j) < min(to_merge[0], to_merge[1]):
+                        # Break ties by choosing pair with lowest min id
+                        new_modularity = trial_modularity
+                        to_merge = (i, j, new_modularity - old_modularity)
+                # Un-merge
+                trial_communities[i] = u
+                trial_communities[j] = v
+        if to_merge is not None:
+            # If the best merge improves modularity, use it
+            merges.append(to_merge)
+            i, j, dq = to_merge
+            u, v = communities[i], communities[j]
+            communities[j] = u | v
+            communities[i] = frozenset([])
+    # Remove empty communities and sort
+    return sorted((c for c in communities if len(c) > 0), key=len, reverse=True)