def centrality_leaders(budgetYears):
network = load_network_for(budgetYears)
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
topK = 10
candidates, rankings = cl.centrality_leaders(g)
ordered_list = []
for r in range(len(rankings))[:topK]:
#logger.info('tier: %d'%r)
for i in list(rankings[r]):
node_name = g.vs[candidates[i]]['name']
ordered_list.append(node_name)
# set the node's centrality_leader attribute, the higher the better
g.vs[candidates[i]]['centrality_leader'] = topK + 1 - r
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
filename = '%s/figures/%s-%s-centrality-leaders.png' % (root_folder(), startBudgetYear, endBudgetYear)
draw(g, filename)
logger.info(ordered_list)
def centrality_leaders(budgetYears):
network = load_network_for(budgetYears)
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
topK = 10
candidates, rankings = cl.centrality_leaders(g)
ordered_list = []
for r in range(len(rankings))[:topK]:
#logger.info('tier: %d'%r)
for i in list(rankings[r]):
node_name = g.vs[candidates[i]]['name']
ordered_list.append(node_name)
# set the node's centrality_leader attribute, the higher the better
g.vs[candidates[i]]['centrality_leader'] = topK + 1 - r
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
filename = '%s/figures/%s-%s-centrality-leaders.png' % (
root_folder(), startBudgetYear, endBudgetYear)
draw(g, filename)
logger.info(ordered_list)
def update_graphml(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
network = load_network_for(budgetYears)
network.g.vs['centrality_leader'] = 0
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
topK = 50
candidates, rankings = cl.centrality_leaders(g)
#ordered_list = []
for r in range(len(rankings))[:topK]:
#logger.info('tier: %d'%r)
for i in list(rankings[r]):
node_name = g.vs[candidates[i]]['name']
# ordered_list.append(node_name)
# set the node's centrality_leader attribute, the higher the better
#g.vs[candidates[i]]['centrality_leader'] = topK + 1 - r
node = network.g.vs.select(name_eq=node_name)
node['centrality_leader'] = topK - r
#logger.info(topK - r)
# logger.info(node['name'])
filename = '%s/data/networks/%d-%d.graphml' % (root_folder(),
startBudgetYear, endBudgetYear)
network.write(filename)
def load_network_for(budgetYears): startBudgetYear = budgetYears[0] endBudgetYear = budgetYears[-1] filename = '%s/data/networks/%d-%d.graphml'%(root_folder(),startBudgetYear, endBudgetYear) network = ResearchCollaborationNetwork.read(filename) return network
def load_network_for(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
filename = '%s/data/networks/%d-%d.graphml' % (
root_folder(), startBudgetYear, endBudgetYear)
network = ResearchCollaborationNetwork.read(filename)
return network
def network_to_d3(budgetYears):
network = load_network_for(budgetYears)
#network = ResearchCollaborationNetwork.read(budgetYears)
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
filename = '%s/data/networks/%s-%s-complete.json' % (
root_folder(), startBudgetYear, endBudgetYear)
ResearchCollaborationNetwork.d3(network.g, filename)
# remove isolated nodes
g = network.g.copy()
g = ResearchCollaborationNetwork.simplify(g)
filename = '%s/data/networks/%s-%s.json' % (root_folder(), startBudgetYear,
endBudgetYear)
ResearchCollaborationNetwork.d3(g, filename)
# only the largest components
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
filename = '%s/data/networks/%s-%s-largest-component.json' % (
root_folder(), startBudgetYear, endBudgetYear)
ResearchCollaborationNetwork.d3(g, filename)
def draw_g(budgetYears): network = load_network_for(budgetYears) g = network.g.copy() #g = g.simplify(multiple=True, loops=True,combine_edges=sum) # convert to undirected #g.to_undirected(combine_edges=sum) g = ResearchCollaborationNetwork.simplify(g) startBudgetYear = budgetYears[0] endBudgetYear = budgetYears[-1] filename = '%s/figures/%s-%s-%d.png'%(root_folder(),startBudgetYear, endBudgetYear,len(g.vs)) #logger.info(g.summary()) draw(g, filename) gl = ResearchCollaborationNetwork.largest_component(g) filename = '%s/figures/%s-%s-%d-largest-component.png'%(root_folder(),startBudgetYear, endBudgetYear,len(gl.vs)) draw(gl, filename)
def network_to_d3(budgetYears):
network = load_network_for(budgetYears)
#network = ResearchCollaborationNetwork.read(budgetYears)
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
filename = '%s/data/networks/%s-%s-complete.json' % (root_folder(),
startBudgetYear, endBudgetYear)
ResearchCollaborationNetwork.d3(network.g, filename)
# remove isolated nodes
g = network.g.copy()
g = ResearchCollaborationNetwork.simplify(g)
filename = '%s/data/networks/%s-%s.json' % (root_folder(),
startBudgetYear, endBudgetYear)
ResearchCollaborationNetwork.d3(g, filename)
# only the largest components
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
filename = '%s/data/networks/%s-%s-largest-component.json' % (root_folder(),
startBudgetYear, endBudgetYear)
ResearchCollaborationNetwork.d3(g, filename)
def test():
from matplotlib import rc
rc('text', usetex=False)
rc('font', family='serif')
fig = plt.figure(figsize=(8, 16))
ax = fig.add_subplot(1, 2, 1)
task = 'per_user'
roc.roc_curve_init(ax)
area, [ax, b] = plot_auc(2006, 2009, task, ax, 'b')
area, [ax, g] = plot_auc(2010, 2012, task, ax, 'g')
area, [ax, r] = plot_auc(2006, 2012, task, ax, 'r')
ax.set_title('Per-user Model')
#f.subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=.3, hspace=.2)
#plt.legend(loc=4)
#plt.savefig('%s/figures/%s-roc-curve.eps'%(root_folder(), task),bbox_inches='tight', dpi=600)
#plt.close()
task = 'per_network'
ax = fig.add_subplot(1, 2, 2)
roc.roc_curve_init(ax)
area, [ax, b] = plot_auc(2006, 2009, task, ax, 'b')
area, [ax, g] = plot_auc(2010, 2012, task, ax, 'g')
area, [ax, r] = plot_auc(2006, 2012, task, ax, 'r')
ax.set_title('Per-network Model')
fig.subplots_adjust(left=None,
bottom=None,
right=None,
top=None,
wspace=.3,
hspace=.2)
#plt.figlegend((b, g, r), ('RCN (2006 - 2009)', 'RCN (2010 - 2012)', 'RCN (2006 - 2012)'), 'center')
#plt.legend(loc='lower center', bbox_to_anchor=(0.5, -0.05), ncol=5, fancybox=True)
plt.legend((b, g, r),
('RCN (2006 - 2009)', 'RCN (2010 - 2012)', 'RCN (2006 - 2012)'),
bbox_to_anchor=(-0.15, -0.4),
loc='center',
prop={'size': 12})
plt.savefig('%s/figures/roc-curve.eps' % (root_folder()),
bbox_inches='tight',
dpi=600)
def plot_auc(startBudgetYear,endBudgetYear, task, ax, color): filename = '%s/data/%s-%s.%s.roc.samples.npy'%(root_folder(),startBudgetYear, endBudgetYear, task) roc_samples = np.load(filename) labels = [] scores = [] for k, label, score in roc_samples: labels.append(np.float(label)) scores.append(np.float(score)) area, [ax, lines] = roc.roc_curve(labels=np.array(labels),scores=np.array(scores), ax=ax, linewidth=1.5, color=color) return area, [ax, lines]
def plot_all_combined():
import matplotlib.pyplot as plt
budgetYears = range(2006,2010)
logger.info("================================================================")
logger.info(budgetYears)
wg_rcn_2006_2009, degree_rcn_2006_2009, strength_rcn_2006_2009 = get_data(budgetYears)
budgetYears = range(2010,2013)
logger.info("================================================================")
logger.info(budgetYears)
wg_rcn_2010_2012, degree_rcn_2010_2012, strength_rcn_2010_2012 = get_data(budgetYears)
budgetYears = range(2006,2013)
logger.info("================================================================")
logger.info(budgetYears)
wg_rcn_2006_2012, degree_rcn_2006_2012, strength_rcn_2006_2012 = get_data(budgetYears)
f = plt.figure(figsize=(12,6))
data = strength_rcn_2006_2009
data_inst = 1
units = 'RCN 2006 - 2009'
plot_powerlaw_combined(data, data_inst, f, units)
data_inst = 2
data = strength_rcn_2010_2012
units = 'RCN 2010 - 2012'
plot_powerlaw_combined(data, data_inst, f, units)
data_inst = 3
data = strength_rcn_2006_2012
units = 'RCN 2006 - 2012'
plot_powerlaw_combined(data, data_inst, f, units)
f.subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=.3, hspace=.2)
f.savefig('%s/figures/powerlaw_degree_distribution.eps'%(root_folder()), bbox_inches='tight')
def rwr_scores(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
logger.info('---------------- %s-%s -------------------' %
(startBudgetYear, endBudgetYear))
network = load_network_for(budgetYears)
#network = ResearchCollaborationNetwork.read(budgetYears)
g = network.g.copy()
ResearchCollaborationNetwork.simplify(g)
logger.info(g.summary())
adj = np.array(g.get_adjacency(igraph.GET_ADJACENCY_BOTH).data)
links = []
m = len(g.vs)
for i in range(m):
for j in range(i + 1, m):
key = '%d,%d' % (i, j)
links.append(key)
rwr_scores = pgrank.rwr_score(g, links)
rwrs = {}
for link, score in rwr_scores.items():
v = link.split(',')
v1 = int(v[0])
v2 = int(v[1])
key = '%s,%s' % (g.vs[v1]['name'], g.vs[v2]['name'])
if(float(score) > 0.001):
rwrs[key] = score
filename = '%s/data/networks/%d-%d-rwr.json' % (root_folder(),
startBudgetYear, endBudgetYear)
with open(filename, 'w') as out:
json.dump(rwrs, out)
def rwr_scores(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
logger.info('---------------- %s-%s -------------------' %
(startBudgetYear, endBudgetYear))
network = load_network_for(budgetYears)
#network = ResearchCollaborationNetwork.read(budgetYears)
g = network.g.copy()
ResearchCollaborationNetwork.simplify(g)
logger.info(g.summary())
adj = np.array(g.get_adjacency(igraph.GET_ADJACENCY_BOTH).data)
links = []
m = len(g.vs)
for i in range(m):
for j in range(i + 1, m):
key = '%d,%d' % (i, j)
links.append(key)
rwr_scores = pgrank.rwr_score(g, links)
rwrs = {}
for link, score in rwr_scores.items():
v = link.split(',')
v1 = int(v[0])
v2 = int(v[1])
key = '%s,%s' % (g.vs[v1]['name'], g.vs[v2]['name'])
if (float(score) > 0.001):
rwrs[key] = score
filename = '%s/data/networks/%d-%d-rwr.json' % (
root_folder(), startBudgetYear, endBudgetYear)
with open(filename, 'w') as out:
json.dump(rwrs, out)
def plot_auc(startBudgetYear, endBudgetYear, task, ax, color):
filename = '%s/data/%s-%s.%s.roc.samples.npy' % (
root_folder(), startBudgetYear, endBudgetYear, task)
roc_samples = np.load(filename)
labels = []
scores = []
for k, label, score in roc_samples:
labels.append(np.float(label))
scores.append(np.float(score))
area, [ax, lines] = roc.roc_curve(labels=np.array(labels),
scores=np.array(scores),
ax=ax,
linewidth=1.5,
color=color)
return area, [ax, lines]
def test():
from matplotlib import rc
rc('text', usetex=False)
rc('font', family='serif')
fig = plt.figure(figsize=(8,16))
ax = fig.add_subplot(1,2,1)
task = 'per_user'
roc.roc_curve_init(ax)
area, [ax, b] = plot_auc(2006, 2009, task, ax, 'b')
area, [ax, g] = plot_auc(2010, 2012, task, ax, 'g')
area, [ax, r] = plot_auc(2006, 2012, task, ax, 'r')
ax.set_title('Per-user Model')
#f.subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=.3, hspace=.2)
#plt.legend(loc=4)
#plt.savefig('%s/figures/%s-roc-curve.eps'%(root_folder(), task),bbox_inches='tight', dpi=600)
#plt.close()
task = 'per_network'
ax = fig.add_subplot(1,2,2)
roc.roc_curve_init(ax)
area, [ax, b] = plot_auc(2006, 2009, task, ax, 'b')
area, [ax, g] = plot_auc(2010, 2012, task, ax, 'g')
area, [ax, r] = plot_auc(2006, 2012, task, ax, 'r')
ax.set_title('Per-network Model')
fig.subplots_adjust(left=None, bottom=None, right=None, top=None, wspace=.3, hspace=.2)
#plt.figlegend((b, g, r), ('RCN (2006 - 2009)', 'RCN (2010 - 2012)', 'RCN (2006 - 2012)'), 'center')
#plt.legend(loc='lower center', bbox_to_anchor=(0.5, -0.05), ncol=5, fancybox=True)
plt.legend((b, g, r), ('RCN (2006 - 2009)', 'RCN (2010 - 2012)', 'RCN (2006 - 2012)'), bbox_to_anchor=(-0.15,-0.4),loc='center', prop={'size':12})
plt.savefig('%s/figures/roc-curve.eps'%(root_folder()),bbox_inches='tight', dpi=600)
def update_graphml(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
network = load_network_for(budgetYears)
network.g.vs['centrality_leader'] = 0
g = network.g.copy()
g = ResearchCollaborationNetwork.largest_component(g)
topK = 50
candidates, rankings = cl.centrality_leaders(g)
#logger.info(candidates)
#logger.info(rankings)
#ordered_list = []
for r in range(len(rankings))[:topK]:
logger.info('tier: %d' % r)
for i in list(rankings[r]):
node_name = g.vs[candidates[i]]['name']
# ordered_list.append(node_name)
# set the node's centrality_leader attribute, the higher the better
#g.vs[candidates[i]]['centrality_leader'] = topK + 1 - r
node = network.g.vs.select(name_eq=node_name)
#logger.info(node['name'])
node['centrality_leader'] = r + 1
#logger.info(topK - r)
# logger.info(node['name'])
filename = '%s/data/networks/%d-%d.graphml' % (
root_folder(), startBudgetYear, endBudgetYear)
network.write(filename)
def per_network(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
logger.info('---------------- %s-%s -------------------'%(startBudgetYear, endBudgetYear))
network = load_network_for(budgetYears)
g = network.g.copy()
ResearchCollaborationNetwork.simplify(g)
logger.info(g.summary())
# randomly pick 20 users
candidates = range(len(g.vs))
shuffle(candidates)
candidates = candidates[:20]
adj = np.array(g.get_adjacency(igraph.GET_ADJACENCY_BOTH).data)
m, _ = adj.shape
nonobservedlinks = {}
nonobserved_actual_edges = []
nonobserved_nonexist_edges = []
for i in range(m):
# undirectd graph, so only care if the source is in candidates or not
if i not in candidates:
continue
for j in range(i + 1, m):
key = '%d,%d'%(i,j)
nonobservedlinks[key] = adj[i,j]
if adj[i,j] > 0:
nonobserved_actual_edges.append(key)
else:
nonobserved_nonexist_edges.append(key)
#logger.info('-----original graph:-----\r\n %s \r\n -----end original graph:-----'%g.summary())
auc = 0.0
apk = {3: 0.0, 5: 0.0, 10: 0.0}
kfold = 10
cnt = 0;
roc_samples = []
for ((es_p_training, es_p_validation), (es_m_training, es_m_validation)) in zip(utils.k_fold_cross_validation(list(nonobserved_actual_edges), kfold), utils.k_fold_cross_validation(list(nonobserved_nonexist_edges), kfold)):
logger.info('--------iteration %d-------------'%cnt)
logger.info('xxxxxxxxxxxxxxxxxxxxxxxx')
logger.info('positive training: %d'%len(es_p_training))
logger.info('positive validation: %d'%len(es_p_validation))
logger.info('------------------------')
logger.info('negative training: %d'%len(es_m_training))
logger.info('negative validation: %d'%len(es_m_validation))
#logger.info('xxxxxxxxxxxxxxxxxxxxxxxx')
training = es_p_training + es_m_training
validation = es_p_validation + es_m_validation
#logger.info('training: %d; valiation: %d'%(len(training), len(validation)))
# create training graph
trainingG = g.copy()
edges_2_delete = []
#// remove edges from the validation set
for link in validation:
v = link.split(',')
v1 = int(v[0])
v2 = int(v[1])
eId = trainingG.get_eid(v1,v2, directed=False, error=False)
if eId != -1:
edges_2_delete.append(eId)
trainingG.delete_edges(edges_2_delete)
#logger.info('-----training graph:-----\r\n %s \r\n -----end training graph:-----'%trainingG.summary())
rwr_scores = pgrank.rwr_score(trainingG, validation)
actual = []
posterior = []
actual_edges = []
for k in validation:
actual.append(nonobservedlinks[k])
if nonobservedlinks[k] > 0:
actual_edges.append(k)
posterior.append(rwr_scores[k])
roc_samples.append((k, nonobservedlinks[k], rwr_scores[k]))
#logger.info('actual edges: %s'%actual_edges)
#logger.info('posterior: %s'%posterior)
auc_ = benchmarks.auc(actual, posterior)
auc += auc_
#area, [ax, lines] = roc.roc_curve(labels=np.array(actual),scores=np.array(posterior))
for topK, p in apk.iteritems():
predictedIndexes = sorted(range(len(posterior)), reverse=True, key=lambda k: posterior[k])[:topK]
predicted = np.array(validation)[predictedIndexes]
apk_ = benchmarks.apk(actual_edges, predicted, topK)
apk[topK] += apk_
cnt += 1
# take a look at http://www.machinedlearnings.com/2012/06/thought-on-link-prediction.html
logger.info('auc: %f'%(auc/kfold))
for topK, p in apk.iteritems():
logger.info('ap@%d: %f'%(topK, (apk[topK]/kfold)))
#plt.show()
np.save('%s/data/%s-%s.per_network.roc.samples.npy'%(root_folder(),startBudgetYear, endBudgetYear), np.array(roc_samples))
def per_candidate(budgetYears):
startBudgetYear = budgetYears[0]
endBudgetYear = budgetYears[-1]
logger.info('---------------- %s-%s -------------------'%(startBudgetYear, endBudgetYear))
network = load_network_for(budgetYears)
g = network.g.copy()
ResearchCollaborationNetwork.simplify(g)
logger.info(g.summary())
adj = np.array(g.get_adjacency(igraph.GET_ADJACENCY_BOTH).data)
m, _ = adj.shape
cNodes = g.vs.select(_degree_gt=15) #range(len(g.vs))
candidates = []
for cNode in cNodes:
candidates.append(cNode.index)
shuffle(candidates)
candidates = candidates[:10]
total_auc = 0.0
precision_at_k = {3: 0.0, 5: 0.0, 10: 0.0}
mapk = precision_at_k
kfold = 5
roc_samples = []
progress = len(candidates)
# for each candidate we do training and testing...
for c in candidates:
logger.info('%d-----------------------'%progress)
nonobservedlinks = {}
nonobserved_actual_edges = []
nonobserved_nonexist_edges = []
# undirectd graph, so only care if the source is in candidates or not
for j in range(m):
key = '%d,%d'%(c,j)
nonobservedlinks[key] = adj[c,j]
#logger.info(adj[c,j])
if adj[c,j] > 0:
nonobserved_actual_edges.append(key)
else:
nonobserved_nonexist_edges.append(key)
cnt = 0
auc = 0.0
#average precision at k is defined per candidate
apk = precision_at_k
for ((es_p_training, es_p_validation), (es_m_training, es_m_validation)) in zip(utils.k_fold_cross_validation(list(nonobserved_actual_edges), kfold), utils.k_fold_cross_validation(list(nonobserved_nonexist_edges), kfold)):
#logger.info('--------iteration %d-------------'%cnt)
#logger.info('xxxxxxxxxxxxxxxxxxxxxxxx')
#logger.info('positive training: %d'%len(es_p_training))
#logger.info('positive validation: %d'%len(es_p_validation))
#logger.info('------------------------')
#logger.info('negative training: %d'%len(es_m_training))
#logger.info('negative validation: %d'%len(es_m_validation))
#logger.info('xxxxxxxxxxxxxxxxxxxxxxxx')
training = es_p_training + es_m_training
validation = es_p_validation + es_m_validation
#logger.info('training: %d; valiation: %d'%(len(training), len(validation)))
# create training graph
trainingG = g.copy()
edges_2_delete = []
#// remove edges from the validation set
for link in validation:
v = link.split(',')
v1 = int(v[0])
v2 = int(v[1])
eId = trainingG.get_eid(v1,v2, directed=False, error=False)
if eId != -1:
edges_2_delete.append(eId)
trainingG.delete_edges(edges_2_delete)
#logger.info('-----training graph:-----\r\n %s \r\n -----end training graph:-----'%trainingG.summary())
rwr_scores = pgrank.rwr_score(trainingG, validation)
for k, rwr_score in rwr_scores.iteritems():
if rwr_score > 1:
logger.info('overflow? rwr_score: %0.2f'%(rwr_score))
actual = []
posterior = []
actual_edges = []
for k in validation:
actual.append(nonobservedlinks[k])
if nonobservedlinks[k] > 0:
actual_edges.append(k)
posterior.append(rwr_scores[k])
roc_samples.append((k, nonobservedlinks[k], rwr_scores[k]))
#logger.info('actual edges: %s'%actual_edges)
#logger.info('posterior: %s'%posterior)
auc_ = benchmarks.auc(actual, posterior)
auc += auc_
total_auc += auc_
#area, [ax, lines] = roc.roc_curve(labels=np.array(actual),scores=np.array(posterior))
for topK, p in mapk.iteritems():
predictedIndexes = sorted(range(len(posterior)), reverse=True, key=lambda k: posterior[k])[:topK]
predicted = np.array(validation)[predictedIndexes]
apk_ = benchmarks.apk(actual_edges, predicted, topK)
apk[topK] += apk_
mapk[topK] += apk_
cnt += 1
logger.info('%d: auc: %f'%(c, float(auc)/kfold))
for topK, p in apk.iteritems():
logger.info('%d: ap@%d: %f'%(c, topK, (apk[topK]/kfold)))
progress -= 1
logger.info('auc: %f'%(float(total_auc)/(kfold*len(candidates))))
for topK, p in mapk.iteritems():
logger.info('map@%d: %f'%(topK, (mapk[topK]/(kfold*len(candidates)))))
np.save('%s/data/%s-%s.per_user.roc.samples.npy'%(root_folder(),startBudgetYear, endBudgetYear), np.array(roc_samples))