def evaluate_embeddings(embeddings,
edges,
cda=True,
greedy_routing=False,
cda_max_vertices=1000,
gr_max_pairs=10000,
eval_core=False,
core_exponent=0.5):
"evaluate quality of embeddings compared with real elge set"
report = []
# get connected component
true_graph = nx.Graph()
true_graph.add_edges_from(edges)
Gcc = sorted(nx.connected_component_subgraphs(true_graph),
key=len,
reverse=True)
true_graph = Gcc[0]
# use BallTree for efficient graph construction
print "construct BallTree"
vertices = list(true_graph.nodes())
edges = list(true_graph.edges())
n = len(vertices)
if eval_core:
vertices = [
v for v in vertices if true_graph.degree(v) >= n**core_exponent
]
true_graph = true_graph.subgraph(vertices)
edges = list(true_graph.edges())
embeddings_array = np.array([embeddings[v] for v in vertices])
bt = BallTree(embeddings_array, metric=distance)
degrees = defaultdict(int)
print "compute number of correct directed arcs"
for v1, v2 in edges:
degrees[v1] += 1
degrees[v2] += 1
# compute number of correct DIRECTED arcs assuming that degrees are known
if cda:
all_correct_arcs = set()
cda_vertices = vertices[:]
if len(cda_vertices) > cda_max_vertices:
np.random.shuffle(cda_vertices)
cda_vertices = cda_vertices[:cda_max_vertices]
for v_i, v in enumerate(cda_vertices):
start = time.time()
degree = degrees[v]
dist, ind = bt.query(np.array(embeddings[v]).reshape(1, -1),
k=degree +
1) # one of neighbors is vertex inself
neigh = [vertices[i] for i in ind[0].tolist() if vertices[i] != v]
for ne in neigh:
if make_edge(v, ne) in edges:
all_correct_arcs.add((v, ne))
finish = time.time()
#print "DEBUG: {} / {}, time={}s".format(v_i + 1, len(cda_vertices), datetime.timedelta(seconds=finish-start))
report.append([
'ratio of correct arcs for known degrees',
float(len(all_correct_arcs)) / (2 * len(edges))
])
if greedy_routing:
print "compute greedy routing efficiency"
random_pairs = set()
if n * (n - 1) / 2 <= gr_max_pairs:
random_pairs = set(combinations(vertices, 2))
else:
while (len(random_pairs) < gr_max_pairs):
v1 = np.random.choice(vertices)
v2 = np.random.choice(vertices)
if v1 != v2:
random_pairs.add((v1, v2))
total_distribution = defaultdict(int)
success_distribution = defaultdict(int)
complete_fails_distribution = defaultdict(int)
all_path_length_pairs = defaultdict(int)
for i, pair in enumerate(random_pairs):
src, dst = pair
# best path
best_path_length = nx.shortest_path_length(true_graph,
source=src,
target=dst)
total_distribution[0] += 1
total_distribution[best_path_length] += 1
# greedy path
curr_src = src
path_length = 0
seen = set()
while curr_src != dst:
seen.add(curr_src)
# find neighbor closest to destination
unseen_neighbors = filter(lambda x: x not in seen,
true_graph.neighbors(curr_src))
if not len(unseen_neighbors):
# greedy algorithm stuck in 'leaf'
path_length = np.nan
break
def curr_distance(v):
return distance(embeddings[dst], embeddings[v])
closest_neigh = min(unseen_neighbors, key=curr_distance)
path_length += 1
curr_src = closest_neigh
if path_length == best_path_length:
success_distribution[0] += 1
success_distribution[best_path_length] += 1
if np.isnan(path_length):
complete_fails_distribution[0] += 1
complete_fails_distribution[best_path_length] += 1
all_path_length_pairs[(best_path_length, path_length)] += 1
all_success = success_distribution[0]
all_complete_fails = complete_fails_distribution[0]
all_total = total_distribution[0]
all_ratio = float(all_success) / all_total * 100
print "Complete fails: {} / {} ({:.2f} %)".format(
all_complete_fails, all_total,
float(all_complete_fails) / all_total * 100)
print "Success: {} / {} ({:.2f} %)".format(all_success, all_total,
all_ratio)
for pl in sorted(
set(total_distribution.keys())
| set(success_distribution.keys())):
if pl == 0:
continue
total = total_distribution.get(pl, 0)
success = success_distribution.get(pl, 0)
ratio = float(success) / total * 100
print "Success at path length = {}: {} / {} ({:.2f} %)".format(
pl, success, total, ratio)
if False:
# depends on R, bad for subgraphs -- not used
n = len(vertices)
R = 2 * np.log(n)
coshR = np.cosh(R)
predicted_edges = set()
print "predict edges"
for v in vertices:
coords = embeddings[v]
neigh_idx = bt.query_radius(coords, R)
neigh = [
vertices[i] for i in neigh_idx[0].tolist() if vertices[i] != v
]
predicted_edges.update([make_edge(v, ne) for ne in neigh])
report.append(['total_predicted_edges', len(predicted_edges)])
# contingency matrix
print "compute contingency matrix"
report.append(['true positive', len(edges & predicted_edges)])
report.append(['false positive', len(predicted_edges - edges)])
report.append(['false negative', len(edges - predicted_edges)])
report.append(
['true negative', n * (n - 1) / 2 - len(edges | predicted_edges)])
return report
def evaluate_embeddings(embeddings, edges, cda=True, greedy_routing=False, cda_max_vertices=1000, gr_max_pairs=10000):
"evaluate quality of embeddings compared with real elge set"
report = []
# get connected component
true_graph = nx.Graph()
true_graph.add_edges_from(edges)
Gcc=sorted(nx.connected_component_subgraphs(true_graph), key=len, reverse=True)
true_graph=Gcc[0]
# use BallTree for efficient graph construction
print "construct BallTree"
vertices = list(true_graph.nodes())
n = len(vertices)
embeddings_array = np.array([embeddings[v] for v in vertices])
bt = BallTree(embeddings_array, metric=distance)
degrees = defaultdict(int)
print "compute number of correct directed arcs"
for v1, v2 in edges:
degrees[v1] += 1
degrees[v2] += 1
# compute number of correct DIRECTED arcs assuming that degrees are known
if cda:
all_correct_arcs = set()
cda_vertices = vertices[:]
if len(cda_vertices) > cda_max_vertices:
np.random.shuffle(cda_vertices)
cda_vertices = cda_vertices[:cda_max_vertices]
for v_i, v in enumerate(cda_vertices):
start = time.time()
degree = degrees[v]
dist, ind = bt.query(embeddings[v], k=degree+1) # one of neighbors is vertex inself
neigh = [vertices[i] for i in ind[0].tolist() if vertices[i] != v]
for ne in neigh:
if make_edge(v, ne) in edges:
all_correct_arcs.add((v, ne))
finish = time.time()
#print "DEBUG: {} / {}, time={}s".format(v_i + 1, len(cda_vertices), datetime.timedelta(seconds=finish-start))
report.append(['ratio of correct arcs for known degrees', float(len(all_correct_arcs)) / (2 * len(edges))])
if greedy_routing:
print "compute greedy routing efficiency"
random_pairs = set()
if n * (n-1) / 2 <= gr_max_pairs:
random_pairs = set(combinations(vertices, 2))
else:
while(len(random_pairs) < gr_max_pairs):
v1 = np.random.choice(vertices)
v2 = np.random.choice(vertices)
if v1 != v2:
random_pairs.add((v1, v2))
total_distribution = defaultdict(int)
success_distribution = defaultdict(int)
for i, pair in enumerate(random_pairs):
src, dst = pair
# best path
best_path_length = nx.shortest_path_length(true_graph, source=src, target=dst)
total_distribution[0] += 1
total_distribution[best_path_length] += 1
# greedy path
curr_src = src
path_length = 0
seen = set()
while curr_src != dst:
seen.add(curr_src)
# find neighbor closest to destination
unseen_neighbors = filter(lambda x: x not in seen, true_graph.neighbors(curr_src))
if not len(unseen_neighbors):
# greedy algorithm stuck in 'leaf'
path_length = np.nan
break
def curr_distance(v):
return distance(embeddings[dst], embeddings[v])
closest_neigh = min(unseen_neighbors, key=curr_distance)
path_length += 1
curr_src = closest_neigh
if path_length == best_path_length:
success_distribution[0] += 1
success_distribution[best_path_length] += 1
all_success = success_distribution[0]
all_total = total_distribution[0]
all_ratio = float(all_success) / all_total * 100
print "Total: {} / {} ({:.2f} %)".format(all_success, all_total, all_ratio)
for pl in sorted(set(total_distribution.keys()) | set(success_distribution.keys())):
if pl == 0:
continue
total = total_distribution.get(pl, 0)
success = success_distribution.get(pl, 0)
ratio = float(success) / total * 100
print "Path length = {}: {} / {} ({:.2f} %)".format(pl, success, total, ratio)
if False:
# depends on R, bad for subgraphs -- not used
n = len(vertices)
R = 2 * np.log(n)
coshR = np.cosh(R)
predicted_edges = set()
print "predict edges"
for v in vertices:
coords = embeddings[v]
neigh_idx = bt.query_radius(coords, R)
neigh = [vertices[i] for i in neigh_idx[0].tolist() if vertices[i] != v]
predicted_edges.update([make_edge(v, ne) for ne in neigh])
report.append(['total_predicted_edges', len(predicted_edges)])
# contingency matrix
print "compute contingency matrix"
report.append(['true positive', len(edges & predicted_edges)])
report.append(['false positive', len(predicted_edges - edges)])
report.append(['false negative', len(edges - predicted_edges)])
report.append(['true negative', n*(n-1)/2 - len(edges | predicted_edges)])
return report
def find_embeddings(
vertices,
edges,
mode,
learning_rate=0.1,
n_epoch=100,
ratio_to_second=2.0,
ratio_between_first=1.0,
ratio_random=1.0,
silent=False,
):
"find (r, phi) for each vertex"
vertices = list(vertices)
n = len(vertices)
R = 2 * np.log(n)
print "mode: {}".format(mode)
np.random.seed(0)
degrees = defaultdict(int)
print "count degrees"
for v1, v2 in edges:
degrees[v1] += 1
degrees[v2] += 1
if mode == "random":
# phi=rand(0, 2pi), r = rand(0,R)
return {v: (np.random.uniform(0.0, R), np.random.uniform(0.0, 2 * np.pi)) for v in vertices}
elif mode == "degrees":
# phi=rand(0,2pi), r = 2log(n/k)
return {v: (2 * np.log(n / degrees[v]), np.random.uniform(0.0, 2 * np.pi)) for v in vertices}
elif mode.startswith("fit"):
x0 = []
for (r, phi) in zip(
[2 * np.log(n / degrees[v]) for v in vertices], [np.random.uniform(0.0, 2 * np.pi) for v in vertices]
):
x0.append(r)
x0.append(phi)
x0 = np.array(x0)
nedges = set()
all_nedges = set()
for (v1, v2) in combinations(vertices, 2):
# if (v1, v2) not in edges and (v2, v1) not in edges:
e = make_edge(v1, v2)
if e not in edges:
all_nedges.add(e)
if mode == "fit_random":
a = list(all_nedges)
random.shuffle(a)
nedges = set(a[: len(edges)])
elif mode == "fit_degrees":
K = float(ratio_to_second) # ratio of nedges to second neighbour
L = float(ratio_between_first) # ratio of nedges between first neighbours
M = float(ratio_random) # ratio of random nedges
# free_nedges = all_nedges.copy()
G = nx.Graph()
G.add_edges_from(edges)
srt_vertices = sorted(degrees.keys(), key=lambda v: -degrees[v])
shuf_vertices = srt_vertices[:]
random.shuffle(shuf_vertices)
for v in srt_vertices:
# get first neighbours
first_neigh = set(G.neighbors(v))
# get second neighbours
second_neigh = set()
for neigh in first_neigh:
second_neigh.update(G.neighbors(neigh))
second_neigh.remove(v)
n_vertex_nedges = 0
# from v to second neighbours
for i, sec_n in enumerate(second_neigh):
# print "i: {}".format(i)
if i + 1 > degrees[v] * K:
continue
e = make_edge(v, sec_n)
if e not in nedges:
nedges.add(e)
n_vertex_nedges += 1
# between first neighbours
for j, pair in enumerate(combinations(first_neigh, 2)):
# print "j: {}".format(j)
if j + 1 > degrees[v] * L:
continue
v1, v2 = pair
e = make_edge(v1, v2)
if e not in nedges:
nedges.add(e)
# random edges
max_n_random_vertices = int(degrees[v] * M)
n_random_vertices = 0
for rand_v in shuf_vertices:
if n_random_vertices >= max_n_random_vertices:
break
e = make_edge(v, rand_v)
if e not in nedges and e not in edges:
nedges.add(e)
n_random_vertices += 1
else:
nedges = all_nedges.copy()
print "number of nedges={}".format(len(nedges))
q = Q(vertices, edges, nedges)
grad_q = GradQ(vertices, edges, nedges)
if mode == "fit_degrees_sgd":
print "Learning rate: {}".format(learning_rate)
print "Ratio to second: {}".format(ratio_to_second)
print "Ratio between first: {}".format(ratio_between_first)
print "Ratio random: {}".format(ratio_random)
G = nx.Graph()
G.add_edges_from(edges)
# construct connected(!) core
core_exponent = 0.4
core_vertices, fringe_vertices = [], []
# one-pass split by condition
for v in vertices:
core_vertices.append(v) if degrees[v] >= n ** core_exponent else fringe_vertices.append(v)
# add vertices to ensure connectivity of core
fringe_vertices.sort(key=lambda v: -degrees[v])
while not nx.is_connected(G.subgraph(core_vertices)):
core_vertices.append(fringe_vertices.pop(0))
print "Core size: {}".format(len(core_vertices))
G_core = G.subgraph(core_vertices)
print "Is core connected:", nx.is_connected(G_core)
# loss_function = MSE(binary_edges=True)
loss_function = LogLoss(binary_edges=True)
optimizer = SGD(n_epoch=n_epoch, learning_rate=learning_rate, verbose=not silent)
FRINGE_FRACTION = 0.1
max_fringe_size = int(G.number_of_nodes() * FRINGE_FRACTION)
curr_graph = G.subgraph(core_vertices)
curr_core_vertices = set(core_vertices)
curr_embedding_model = PoincareModel(curr_graph, fit_radius=False)
curr_pair_generator = BinaryPairGenerator(curr_graph, batch_size=1)
optimizer.optimize_embedding(curr_embedding_model, loss_function, curr_pair_generator)
for i in range(int(1 / FRINGE_FRACTION) + 1):
total_fringe = fringe(G, curr_core_vertices)
# print "DEBUG:", curr_graph. number_of_nodes(), len(curr_core_vertices), len(total_fringe)
fringe_vertices = set(sorted(total_fringe, key=lambda v: -G.degree(v))[:max_fringe_size])
# print "DEBUG:", i+1, fringe_vertices
if not fringe_vertices:
break
curr_graph = G.subgraph(curr_core_vertices | fringe_vertices)
curr_embedding_model = PoincareModel(curr_graph, fit_radius=False, init_embedding=curr_embedding_model)
curr_pair_generator = BinaryPairGenerator(curr_graph, batch_size=1)
optimizer.optimize_embedding(
curr_embedding_model, loss_function, curr_pair_generator, fixed_vertices=curr_core_vertices
)
curr_core_vertices |= fringe_vertices
embedding_model = curr_embedding_model
"""
core_embedding_model = PoincareModel(G_core, fit_radius=False)
core_pair_generator = BinaryPairGenerator(G_core, batch_size=1)
optimizer.optimize_embedding(core_embedding_model, loss_function, core_pair_generator)
#optimizer = SGD(n_epoch=n_epoch, learning_rate=learning_rate, verbose=not silent)
embedding_model = PoincareModel(G, fit_radius=False, init_embedding=core_embedding_model)
pair_generator = BinaryPairGenerator(G, batch_size=1)
optimizer.optimize_embedding(embedding_model, loss_function, pair_generator, fixed_vertices=core_vertices)
#print "Radius before: {}".format(embedding_model.embedding['radius'])
#print "Radius after: {}".format(embedding_model.embedding['radius'])
"""
return (embedding_model.embedding["vertices"], {"core": list(G.edges())})
else:
print "Check gradient: ", check_grad(q, grad_q, x0)
res = minimize(q, x0, method="BFGS", jac=grad_q)
# print res
x = res.x
retval = {}
for i in range(len(vertices)):
r = x[2 * i]
phi = x[2 * i + 1]
retval[vertices[i]] = (r, phi)
return retval
else:
raise Exception("unknown mode")
def find_embeddings(vertices,
edges,
mode,
learning_rate=0.1,
n_epoch=100,
ratio_to_second=2.,
ratio_between_first=1.,
ratio_random=1.,
silent=False):
"find (r, phi) for each vertex"
vertices = list(vertices)
n = len(vertices)
R = 2 * np.log(n)
print "mode: {}".format(mode)
np.random.seed(0)
degrees = defaultdict(int)
print "count degrees"
for v1, v2 in edges:
degrees[v1] += 1
degrees[v2] += 1
if mode == 'random':
# phi=rand(0, 2pi), r = rand(0,R)
return {
v: (np.random.uniform(0.0, R), np.random.uniform(0.0, 2 * np.pi))
for v in vertices
}
elif mode == 'degrees':
# phi=rand(0,2pi), r = 2log(n/k)
return {
v: (2 * np.log(n / degrees[v]), np.random.uniform(0.0, 2 * np.pi))
for v in vertices
}
elif mode.startswith('fit'):
x0 = []
for (r,
phi) in zip([2 * np.log(n / degrees[v]) for v in vertices],
[np.random.uniform(0.0, 2 * np.pi)
for v in vertices]):
x0.append(r)
x0.append(phi)
x0 = np.array(x0)
nedges = set()
all_nedges = set()
for (v1, v2) in combinations(vertices, 2):
#if (v1, v2) not in edges and (v2, v1) not in edges:
e = make_edge(v1, v2)
if e not in edges:
all_nedges.add(e)
if mode == 'fit_random':
a = list(all_nedges)
random.shuffle(a)
nedges = set(a[:len(edges)])
elif mode == 'fit_degrees':
K = float(ratio_to_second) # ratio of nedges to second neighbour
L = float(ratio_between_first
) # ratio of nedges between first neighbours
M = float(ratio_random) # ratio of random nedges
#free_nedges = all_nedges.copy()
G = nx.Graph()
G.add_edges_from(edges)
srt_vertices = sorted(degrees.keys(), key=lambda v: -degrees[v])
shuf_vertices = srt_vertices[:]
random.shuffle(shuf_vertices)
for v in srt_vertices:
# get first neighbours
first_neigh = set(G.neighbors(v))
# get second neighbours
second_neigh = set()
for neigh in first_neigh:
second_neigh.update(G.neighbors(neigh))
second_neigh.remove(v)
n_vertex_nedges = 0
# from v to second neighbours
for i, sec_n in enumerate(second_neigh):
#print "i: {}".format(i)
if i + 1 > degrees[v] * K:
continue
e = make_edge(v, sec_n)
if e not in nedges:
nedges.add(e)
n_vertex_nedges += 1
# between first neighbours
for j, pair in enumerate(combinations(first_neigh, 2)):
#print "j: {}".format(j)
if j + 1 > degrees[v] * L:
continue
v1, v2 = pair
e = make_edge(v1, v2)
if e not in nedges:
nedges.add(e)
# random edges
max_n_random_vertices = int(degrees[v] * M)
n_random_vertices = 0
for rand_v in shuf_vertices:
if n_random_vertices >= max_n_random_vertices:
break
e = make_edge(v, rand_v)
if e not in nedges and e not in edges:
nedges.add(e)
n_random_vertices += 1
else:
nedges = all_nedges.copy()
print "number of nedges={}".format(len(nedges))
q = Q(vertices, edges, nedges)
grad_q = GradQ(vertices, edges, nedges)
if mode == 'fit_degrees_sgd':
print "Learning rate: {}".format(learning_rate)
print "Ratio to second: {}".format(ratio_to_second)
print "Ratio between first: {}".format(ratio_between_first)
print "Ratio random: {}".format(ratio_random)
G = nx.Graph()
G.add_edges_from(edges)
# construct connected(!) core
core_exponent = 0.4
core_vertices, fringe_vertices = [], []
# one-pass split by condition
for v in vertices:
core_vertices.append(v) if degrees[
v] >= n**core_exponent else fringe_vertices.append(v)
# add vertices to ensure connectivity of core
fringe_vertices.sort(key=lambda v: -degrees[v])
while not nx.is_connected(G.subgraph(core_vertices)):
core_vertices.append(fringe_vertices.pop(0))
print "Core size: {}".format(len(core_vertices))
G_core = G.subgraph(core_vertices)
print "Is core connected:", nx.is_connected(G_core)
#loss_function = MSE(binary_edges=True)
loss_function = LogLoss(binary_edges=True)
optimizer = SGD(n_epoch=n_epoch,
learning_rate=learning_rate,
verbose=not silent)
FRINGE_FRACTION = 0.1
max_fringe_size = int(G.number_of_nodes() * FRINGE_FRACTION)
curr_graph = G.subgraph(core_vertices)
curr_core_vertices = set(core_vertices)
curr_embedding_model = PoincareModel(curr_graph, fit_radius=False)
curr_pair_generator = BinaryPairGenerator(curr_graph, batch_size=1)
optimizer.optimize_embedding(curr_embedding_model, loss_function,
curr_pair_generator)
for i in range(int(1 / FRINGE_FRACTION) + 1):
total_fringe = fringe(G, curr_core_vertices)
#print "DEBUG:", curr_graph. number_of_nodes(), len(curr_core_vertices), len(total_fringe)
fringe_vertices = set(
sorted(total_fringe,
key=lambda v: -G.degree(v))[:max_fringe_size])
#print "DEBUG:", i+1, fringe_vertices
if not fringe_vertices:
break
curr_graph = G.subgraph(curr_core_vertices | fringe_vertices)
curr_embedding_model = PoincareModel(
curr_graph,
fit_radius=False,
init_embedding=curr_embedding_model)
curr_pair_generator = BinaryPairGenerator(curr_graph,
batch_size=1)
optimizer.optimize_embedding(curr_embedding_model,
loss_function,
curr_pair_generator,
fixed_vertices=curr_core_vertices)
curr_core_vertices |= fringe_vertices
embedding_model = curr_embedding_model
'''
core_embedding_model = PoincareModel(G_core, fit_radius=False)
core_pair_generator = BinaryPairGenerator(G_core, batch_size=1)
optimizer.optimize_embedding(core_embedding_model, loss_function, core_pair_generator)
#optimizer = SGD(n_epoch=n_epoch, learning_rate=learning_rate, verbose=not silent)
embedding_model = PoincareModel(G, fit_radius=False, init_embedding=core_embedding_model)
pair_generator = BinaryPairGenerator(G, batch_size=1)
optimizer.optimize_embedding(embedding_model, loss_function, pair_generator, fixed_vertices=core_vertices)
#print "Radius before: {}".format(embedding_model.embedding['radius'])
#print "Radius after: {}".format(embedding_model.embedding['radius'])
'''
return (embedding_model.embedding['vertices'], {
'core': list(G.edges())
})
else:
print "Check gradient: ", check_grad(q, grad_q, x0)
res = minimize(q, x0, method='BFGS', jac=grad_q)
#print res
x = res.x
retval = {}
for i in range(len(vertices)):
r = x[2 * i]
phi = x[2 * i + 1]
retval[vertices[i]] = (r, phi)
return retval
else:
raise Exception('unknown mode')