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))
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))