|
a |
|
b/pathaia/graphs/clustering.py |
|
|
1 |
from typing import Any, Dict, Optional, Sequence, Tuple, Union |
|
|
2 |
|
|
|
3 |
import numpy as np |
|
|
4 |
from nptyping import NDArray, Shape |
|
|
5 |
from scipy.sparse import triu |
|
|
6 |
from sortedcontainers import SortedDict |
|
|
7 |
from tqdm import tqdm |
|
|
8 |
|
|
|
9 |
from .object_api import Tree, UGraph |
|
|
10 |
from .types import Edge, Node |
|
|
11 |
|
|
|
12 |
|
|
|
13 |
class AgglomerativeClustering: |
|
|
14 |
r""" |
|
|
15 |
Object used to hierarchically cluster nodes on a graph. Clustering greedily chooses |
|
|
16 |
to merge linked nodes that have minimum distance/strength ratio. Strength between |
|
|
17 |
2 nodes is initially 1 for every edge and 0 when there is no edge, then when 2 nodes |
|
|
18 |
are merged the strength of a newly formed link between the new node and another node |
|
|
19 |
is the weighted (by node population) average of the strengths between the 2 old |
|
|
20 |
nodes and the other node. This algorithm uses centroid linkage clustering (UPGMC). |
|
|
21 |
|
|
|
22 |
Args: |
|
|
23 |
compute_all: whether to initially compute all distances between nodes regardless |
|
|
24 |
of there linkage. |
|
|
25 |
""" |
|
|
26 |
|
|
|
27 |
def __init__(self, compute_all: bool = False): |
|
|
28 |
self.compute_all = compute_all |
|
|
29 |
|
|
|
30 |
def init_graph( |
|
|
31 |
self, |
|
|
32 |
G: UGraph, |
|
|
33 |
feats: Union[Dict[Node, NDArray[Shape["*"], Any]], Sequence[str]], |
|
|
34 |
weights: Optional[Union[Dict[Edge, float], str]] = None, |
|
|
35 |
): |
|
|
36 |
r""" |
|
|
37 |
Initialize main graph attributes (adjacency matrix, n_nodes and features) using |
|
|
38 |
a graph object, a list of features and a list of weights. |
|
|
39 |
|
|
|
40 |
Args: |
|
|
41 |
G: graph to cluster nodes on. |
|
|
42 |
feats: either a dictionary that maps nodes to their corresponding feature |
|
|
43 |
vectors or a sequence of property names that will be used as features. |
|
|
44 |
weights: either a dictionary that maps edges to their corresponding weight |
|
|
45 |
or a property name that will be used as weight. If `None` is passed, |
|
|
46 |
weights are computed using euclidian distances between feature vectors. |
|
|
47 |
""" |
|
|
48 |
self.A = triu(G.A, format="csr").astype(np.float32) |
|
|
49 |
self.n_nodes = G.n_nodes |
|
|
50 |
if isinstance(feats, dict): |
|
|
51 |
feats = [feats[node] for node in G.nodes] |
|
|
52 |
self.feats = np.stack(feats) |
|
|
53 |
else: |
|
|
54 |
self.feats = [] |
|
|
55 |
for node in G.nodes: |
|
|
56 |
self.feats.append([G.nodeprops[feat][node] for feat in feats]) |
|
|
57 |
self.feats = np.array(self.feats) |
|
|
58 |
|
|
|
59 |
if weights is None: |
|
|
60 |
ii, jj = self.A.nonzero() |
|
|
61 |
dists = ((feats[ii] - feats[jj]) ** 2).sum(1) |
|
|
62 |
self.A[ii, jj] = dists |
|
|
63 |
elif isinstance(weights, dict): |
|
|
64 |
for (n1, n2) in weights: |
|
|
65 |
i, j = sorted((G.nodes.index(n1), G.nodes.index(n2))) |
|
|
66 |
self.A[i, j] = weights[n1, n2] ** 2 |
|
|
67 |
else: |
|
|
68 |
for (n1, n2) in G.edges: |
|
|
69 |
i, j = sorted((G.nodes.index(n1), G.nodes.index(n2))) |
|
|
70 |
self.A[i, j] = G.edgeprops[str][(n1, n2)] ** 2 |
|
|
71 |
|
|
|
72 |
def reset(self): |
|
|
73 |
""" |
|
|
74 |
Reset the algorithm attributes. Populations are initiated to 1 for every node, |
|
|
75 |
strengths are initiated to 1 for every edge, dendrogram is emptied. |
|
|
76 |
""" |
|
|
77 |
self.populations_ = {k: 1 for k in range(self.n_nodes)} |
|
|
78 |
ii, jj = self.A.nonzero() |
|
|
79 |
self.centroids_ = {k: self.feats[k] for k in range(self.n_nodes)} |
|
|
80 |
self.links_ = { |
|
|
81 |
k: set(jj[ii == k].tolist() + ii[jj == k].tolist()) |
|
|
82 |
for k in range(self.n_nodes) |
|
|
83 |
} |
|
|
84 |
self.strengths_ = {(i, j): 1 for i, j in zip(ii, jj)} |
|
|
85 |
if self.compute_all: |
|
|
86 |
self.distances_ = { |
|
|
87 |
(i, j): self.distance(i, j) |
|
|
88 |
for i in range(self.n_nodes) |
|
|
89 |
for j in range(i + 1, self.n_nodes) |
|
|
90 |
} |
|
|
91 |
else: |
|
|
92 |
self.distances_ = {(i, j): self.distance(i, j) for i, j in zip(ii, jj)} |
|
|
93 |
self.edges_ = SortedDict( |
|
|
94 |
self.criterion, {(i, j): self.distances_[i, j] for i, j in zip(ii, jj)} |
|
|
95 |
) |
|
|
96 |
self.dendrogram_ = np.zeros((self.n_nodes - 1, 4)) |
|
|
97 |
|
|
|
98 |
def distance(self, i: int, j: int) -> float: |
|
|
99 |
""" |
|
|
100 |
Get squared distance between nodes `i` and `j`. If available in the adjacency |
|
|
101 |
matrix or in the `distance` dictionary it is not recomputed. |
|
|
102 |
|
|
|
103 |
Args: |
|
|
104 |
i: first node. |
|
|
105 |
j: second node. |
|
|
106 |
|
|
|
107 |
Returns: |
|
|
108 |
Squared euclidian distance between i and j. |
|
|
109 |
""" |
|
|
110 |
i, j = sorted((i, j)) |
|
|
111 |
try: |
|
|
112 |
d = self.A[i, j] |
|
|
113 |
except IndexError: |
|
|
114 |
d = self.distances_.get((i, j), 0) |
|
|
115 |
if not d: |
|
|
116 |
d = ((self.centroids_[i] - self.centroids_[j]) ** 2).sum() |
|
|
117 |
return d |
|
|
118 |
|
|
|
119 |
def criterion(self, x: Tuple[int, int]) -> float: |
|
|
120 |
""" |
|
|
121 |
Criterion function used to find the next nodes to merge. Override it to use |
|
|
122 |
another criterion. |
|
|
123 |
|
|
|
124 |
Args: |
|
|
125 |
x: tuple containg the two nodes to merge. |
|
|
126 |
|
|
|
127 |
Returns: |
|
|
128 |
Squared distance between the 2 nodes divided by link strength. |
|
|
129 |
""" |
|
|
130 |
i, j = x |
|
|
131 |
return self.distances_[i, j] / self.strengths_[i, j] |
|
|
132 |
|
|
|
133 |
def create_centroid_link(self, i, j, c, k): |
|
|
134 |
""" |
|
|
135 |
Create a new link between centroid `c` (that comes from merging nodes `i` and |
|
|
136 |
`j`) and node `k`. |
|
|
137 |
|
|
|
138 |
Args: |
|
|
139 |
i: first merged node. |
|
|
140 |
j: second merged node. |
|
|
141 |
c: centroid of nodes `i` and `j`. |
|
|
142 |
k: node linked to either `i` or `j` or both. |
|
|
143 |
""" |
|
|
144 |
if i == k or j == k: |
|
|
145 |
return |
|
|
146 |
ik, ki = sorted((i, k)) |
|
|
147 |
jk, kj = sorted((j, k)) |
|
|
148 |
ck, kc = sorted((c, k)) |
|
|
149 |
ij, ji = sorted((i, j)) |
|
|
150 |
|
|
|
151 |
pi = self.populations_[i] |
|
|
152 |
pj = self.populations_[j] |
|
|
153 |
ri = pi / (pi + pj) |
|
|
154 |
rj = pj / (pi + pj) |
|
|
155 |
try: |
|
|
156 |
dik = self.distances_[ik, ki] |
|
|
157 |
djk = self.distances_[jk, kj] |
|
|
158 |
dij = self.distances_[ij, ji] |
|
|
159 |
dck = ri * dik + rj * djk - ri * rj * dij |
|
|
160 |
except KeyError: |
|
|
161 |
dck = self.distance(c, k) |
|
|
162 |
|
|
|
163 |
self.distances_[ck, kc] = dck |
|
|
164 |
|
|
|
165 |
self.edges_.pop((ik, ki), 0) |
|
|
166 |
self.edges_.pop((jk, kj), 0) |
|
|
167 |
sik = self.strengths_.get((ik, ki), 0) |
|
|
168 |
sjk = self.strengths_.get((jk, kj), 0) |
|
|
169 |
if sik or sjk: |
|
|
170 |
self.strengths_[ck, kc] = ri * sik + rj * sjk |
|
|
171 |
self.edges_[ck, kc] = dck |
|
|
172 |
|
|
|
173 |
self.links_[k].discard(i) |
|
|
174 |
self.links_[k].discard(j) |
|
|
175 |
self.links_[k].add(c) |
|
|
176 |
self.links_[j].discard(k) |
|
|
177 |
self.links_[c].add(k) |
|
|
178 |
|
|
|
179 |
def add_link(self, i: int, j: int): |
|
|
180 |
""" |
|
|
181 |
Create a new link between 2 nodes. |
|
|
182 |
|
|
|
183 |
Args: |
|
|
184 |
i: first node. |
|
|
185 |
j: second node. |
|
|
186 |
""" |
|
|
187 |
i, j = sorted((i, j)) |
|
|
188 |
dij = self.distance(i, j) |
|
|
189 |
self.distances_[i, j] = dij |
|
|
190 |
self.strengths_[i, j] = 1 |
|
|
191 |
self.edges_[i, j] = dij |
|
|
192 |
self.links_[i].add(j) |
|
|
193 |
self.links_[j].add(i) |
|
|
194 |
|
|
|
195 |
def fit( |
|
|
196 |
self, |
|
|
197 |
G: UGraph, |
|
|
198 |
feats: Union[Dict[Node, NDArray[Shape["*"], Any]], Sequence[str]], |
|
|
199 |
weights: Optional[Union[Dict[Edge, float], str]] = None, |
|
|
200 |
): |
|
|
201 |
r""" |
|
|
202 |
Fits on the given graph and completes the dendrogram. A dendrogram is an array |
|
|
203 |
of size :math:`(n-1) \times 4` (whre :math:`n` is the number of nodes) |
|
|
204 |
representing the successive merges of nodes. Each row gives the two merged |
|
|
205 |
nodes, their distance and the size of the resulting cluster. Any new node |
|
|
206 |
resulting from a merge takes the first available index (e.g., the first merge |
|
|
207 |
corresponds to node :math:`n`). |
|
|
208 |
|
|
|
209 |
Args: |
|
|
210 |
G: graph to cluster nodes on. |
|
|
211 |
feats: either a dictionary that maps nodes to their corresponding feature |
|
|
212 |
vectors or a sequence of property names that will be used as features. |
|
|
213 |
weights: either a dictionary that maps edges to their corresponding weight |
|
|
214 |
or a property name that will be used as weight. If `None` is passed, |
|
|
215 |
weights are computed using euclidian distances between feature vectors. |
|
|
216 |
""" |
|
|
217 |
self.init_graph(G, feats, weights) |
|
|
218 |
self.reset() |
|
|
219 |
|
|
|
220 |
c = self.n_nodes |
|
|
221 |
for n in tqdm(range(self.n_nodes - 1), total=self.n_nodes - 1): |
|
|
222 |
if not self.edges_: |
|
|
223 |
cur_dendrogram = self.dendrogram_[:n] |
|
|
224 |
missing = sorted( |
|
|
225 |
[ |
|
|
226 |
k |
|
|
227 |
for k in range(c) |
|
|
228 |
if k not in cur_dendrogram[:, 0] |
|
|
229 |
and k not in cur_dendrogram[:, 1] |
|
|
230 |
] |
|
|
231 |
) |
|
|
232 |
for k, i in enumerate(missing): |
|
|
233 |
for j in missing[k + 1 :]: |
|
|
234 |
self.add_link(i, j) |
|
|
235 |
(i, j), _ = self.edges_.popitem(0) |
|
|
236 |
|
|
|
237 |
pi = self.populations_[i] |
|
|
238 |
pj = self.populations_[j] |
|
|
239 |
ri = pi / (pi + pj) |
|
|
240 |
rj = pj / (pi + pj) |
|
|
241 |
self.dendrogram_[n] = [i, j, self.criterion((i, j)), pi + pj] |
|
|
242 |
|
|
|
243 |
self.centroids_[c] = ri * self.centroids_[i] + rj * self.centroids_[j] |
|
|
244 |
self.populations_[c] = pi + pj |
|
|
245 |
self.links_[c] = set() |
|
|
246 |
|
|
|
247 |
while self.links_[i]: |
|
|
248 |
k = self.links_[i].pop() |
|
|
249 |
self.create_centroid_link(i, j, c, k) |
|
|
250 |
while self.links_[j]: |
|
|
251 |
k = self.links_[j].pop() |
|
|
252 |
self.create_centroid_link(j, i, c, k) |
|
|
253 |
self.links_.pop(i) |
|
|
254 |
self.links_.pop(j) |
|
|
255 |
c += 1 |
|
|
256 |
|
|
|
257 |
def fit_transform( |
|
|
258 |
self, |
|
|
259 |
G: UGraph, |
|
|
260 |
feats: Union[Dict[Node, NDArray[Shape["*"], Any]], Sequence[str]], |
|
|
261 |
weights: Optional[Union[Dict[Edge, float], str]] = None, |
|
|
262 |
) -> Tree: |
|
|
263 |
""" |
|
|
264 |
Fits on the given graph and returns the hierarchical clustering tree. |
|
|
265 |
|
|
|
266 |
Args: |
|
|
267 |
G: graph to cluster nodes on. |
|
|
268 |
feats: either a dictionary that maps nodes to their corresponding feature |
|
|
269 |
vectors or a sequence of property names that will be used as features. |
|
|
270 |
weights: either a dictionary that maps edges to their corresponding weight |
|
|
271 |
or a property name that will be used as weight. If `None` is passed, |
|
|
272 |
weights are computed using euclidian distances between feature vectors. |
|
|
273 |
|
|
|
274 |
Returns: |
|
|
275 |
The tree that describes the hierarchical clustering procedure. |
|
|
276 |
""" |
|
|
277 |
self.fit(G, feats, weights) |
|
|
278 |
|
|
|
279 |
children = [] |
|
|
280 |
parents = [] |
|
|
281 |
nodes = list(range(G.n_nodes)) |
|
|
282 |
|
|
|
283 |
if isinstance(weights, str): |
|
|
284 |
key = weights |
|
|
285 |
else: |
|
|
286 |
key = "weight" |
|
|
287 |
edgeprops = {key: {}} |
|
|
288 |
|
|
|
289 |
if isinstance(feats, dict): |
|
|
290 |
n_feats = next(iter(feats)).shape |
|
|
291 |
nodeprops = { |
|
|
292 |
k: {n: feats[node][k] for n, node in enumerate(G.nodes)} |
|
|
293 |
for k in range(n_feats) |
|
|
294 |
} |
|
|
295 |
else: |
|
|
296 |
nodeprops = { |
|
|
297 |
feat: {n: G.nodeprops[feat][node] for n, node in enumerate(G.nodes)} |
|
|
298 |
for feat in feats |
|
|
299 |
} |
|
|
300 |
nodeprops["population"] = {n: 1 for n in range(G.n_nodes)} |
|
|
301 |
|
|
|
302 |
for k, row in enumerate(self.dendrogram_): |
|
|
303 |
n = k + self.n_nodes |
|
|
304 |
n1, n2 = row[:2] |
|
|
305 |
children[n] = [n1, n2] |
|
|
306 |
parents[n1] = n |
|
|
307 |
parents[n2] = n |
|
|
308 |
nodes.append(n) |
|
|
309 |
if isinstance(feats, dict): |
|
|
310 |
for k, centroid in enumerate(self.centroids_[n]): |
|
|
311 |
nodeprops[k][n] = centroid |
|
|
312 |
else: |
|
|
313 |
for k, feat in enumerate(feats): |
|
|
314 |
nodeprops[feat][n] = self.centroids_[n, k] |
|
|
315 |
edgeprops[key][n, n1] = self.distance(n, n1) ** 0.5 |
|
|
316 |
edgeprops[key][n, n2] = self.distance(n, n2) ** 0.5 |
|
|
317 |
nodeprops["population"][n] = self.populations_[n] |
|
|
318 |
|
|
|
319 |
return Tree(nodes, parents, children, nodeprops, edgeprops) |