def train_models(columns, rf=True, svm=False, logit=False):
# Prepare the data
dataCols = dict(
[(colName,
read_data_col('%s.train.dat' % fileName, valCol=valCol))
for colName, fileName, valCol in COLUMNS
if colName in columns or colName == 'TF']
)
data, colnames = create_matrix(dataCols)
cimt = colnames.index('TF')
y = data[:, cimt]
x = data[:, [c for c in range(len(colnames)) if c != cimt]]
print >> sys.stderr, '\n>> Training model with cols:', colnames
print >> sys.stderr, x
# Fit the models
models = {}
if rf:
rfModel = ensemble.RandomForestClassifier(
n_estimators=100,
## criterion='entropy',
verbose=1,
n_jobs=-1,
## oob_score=True,
## min_samples_leaf=1,
)
rfModel.fit(x, y)
models['rf'] = rfModel
if svm:
svmModel = svm.SVC(kernel='rbf')
svmModel.fit(x, y)
models['svm'] = svmModel
if logit:
logitModel = linear_model.LogisticRegression()
logitModel.fit(x, y)
models['logit'] = logitModel
return {'models' : models, 'colnames' : [c for c in colnames if c!='TF']}
def cross_validation(logit=False):
# Split authors in train/validation
confirmed, deleted = get_train()
aids = confirmed.keys()
shuffle(aids)
cutPoint = int(len(aids) * F)
aidsTrain = aids[:cutPoint]
aidsValid = aids[cutPoint:]
# Prepare the training data
dataCols = dict(
[(colName,
read_data_col('%s.train.dat' % fileName, valCol=valCol))
for colName, fileName, valCol in COLUMNS]
)
data, colnames = create_matrix(dataCols, exclude_aids=aidsValid)
cimt = colnames.index('TF')
ytrain = data[:, cimt]
xtrain = data[:, [c for c in range(len(colnames)) if c != cimt]]
# validation
data, colnames2, pairs = create_matrix(dataCols,
exclude_aids=aidsTrain,
return_pairs=True)
if colnames2 != colnames:
raise WTF
yvalid = data[:, cimt]
xvalid = data[:, [c for c in range(len(colnames)) if c != cimt]]
print >> sys.stderr, '==> %s train; %s valid <==' % (str(xtrain.shape),
str(xvalid.shape))
# Train the model
models = train_models(xtrain, ytrain, logit=logit)
# Calculate performance of each algorithm
performance = {}
for algorithm in models:
print >> sys.stderr, '\n\n--> %s <--\n' % algorithm
print >> sys.stderr, models[algorithm]
# make the predictions
predBin = models[algorithm].predict(xvalid)
pred = models[algorithm].predict_proba(xvalid)
print >> sys.stderr, ''
print >> sys.stderr, yvalid
print >> sys.stderr, predBin
print >> sys.stderr, pred
# extract author-paper scores
decorated = {}
for n in range(len(pairs)):
aid, pid = pairs[n]
score = pred[n, 0]
try:
decorated[aid].append((score, pid, yvalid[n]))
except KeyError:
decorated[aid] = [(score, pid, yvalid[n])]
# get the MAP
MAPs = []
for aid in decorated:
sortedPapers = decorated[aid]
sortedPapers.sort()
ntrue, ntot, MAPTerms = 0, 0, []
for s, p, tf in sortedPapers:
ntot += 1
if tf == 1:
ntrue += 1
MAPTerms.append(float(ntrue)/float(ntot))
MAPs.append(np.mean(MAPTerms))
# output results
performance[algorithm] = np.mean(MAPs)
print >> sys.stderr, '\n>> %s: %f' % (algorithm, np.mean(MAPs))
# Done
print >> sys.stderr, '\n'
return performance
def make_submission():
# Read validation file
validation = get_valid()
# Read scores for validation author-paper pairs
validDataCols = dict(
[(colName,
read_data_col('%s.valid.dat' % fileName, valCol=valCol))
for colName, fileName, valCol in COLUMNS
if colName != 'TF']
)
# Train the models and prepare the validation sets
models, xValid, pairsAP = {}, {}, {}
for aid in validation:
for pid in validation[aid]:
# the available columns for this author-paper pair
try:
cols = [col
for col in validDataCols
if validDataCols[col].has_key(aid)
and validDataCols[col][aid].has_key(pid)]
except:
print aid, pid, aid in validDataCols[col].keys()
raise ValueError
cols.sort()
# train the model if necessary (i.e. if this is the first
# time we see exactly these columns)
modelName = '_'.join(cols)
if modelName not in models:
models[modelName] = train_models(cols)
xValid[modelName] = []
pairsAP[modelName] = []
# validation columns
xValid[modelName].append(
np.array([validDataCols[col][aid][pid]
for col in models[modelName]['colnames']])
)
pairsAP[modelName].append((aid, pid))
# Make the predictions
decorated = {}
for modelName in models:
print >> sys.stderr, '\n>> Making predictions for cols:', modelName
# make all predictions for this model
print >> sys.stderr, np.array(xValid[modelName])
print >> sys.stderr, np.array(xValid[modelName]).shape
pred = models[modelName]['models']['rf'].predict_proba(
np.array(xValid[modelName])
)
print >> sys.stderr, pred
# extract author-paper scores
for n in range(len(pairsAP[modelName])):
aid, pid = pairsAP[modelName][n]
score = pred[n, 0]
try:
decorated[aid].append((score, pid))
except KeyError:
decorated[aid] = [(score, pid)]
# Create the file
outf = open('submit_%s.csv' % strftime("%Y%m%d_%H:%M:%S", localtime()), 'w')
print >> outf, 'AuthorId,PaperIds'
for aid in decorated:
sortedPapers = decorated[aid]
sortedPapers.sort()
print >> outf, \
str(aid) + ',' + ' '.join([str(pid) for s, pid in sortedPapers])
outf.close()
# Create the info file
outf = open('submit_%s.info' % strftime("%Y%m%d_%H:%M:%S", localtime()), 'w')
print >> outf, '\n'.join([str(c) for c in COLUMNS])
outf.close()
('nconference', '../nvenue.train.dat', 4),
('name', '../name.train.dat', 3),
## ('nameinit', '../name.train.dat', 4),
# ('nname', '../nname.train.dat', 3),
('npapers', '../npapers.train.dat', 3),
('nauthors', '../nauthors.train.dat', 3),
# ('coauthors', '../coauthors_diff.train.dat', 3),
## ('zcoauthors', '../coauthors_diff.train.dat', 4),
('affiliation', '../affiliation.train.dat', 3),
('year', '../year.train.dat', 3),
('nvalidated', '../nvalidated.train.dat', 3),
('sumcoauthors', '../sumcoauthors.train.dat', 3),
)
if __name__ == '__main__':
dataCols = dict([(colName, read_data_col(fileName, valCol=valCol))
for colName, fileName, valCol in COLUMNS])
data, colnames = create_matrix(dataCols)
print colnames
cimt = colnames.index('TF')
y = data[:, cimt]
x = data[:, [c for c in range(len(colnames)) if c != cimt]]
N = len(y)
ntrain = int(N * F)
ttindices = range(N)
shuffle(ttindices)
yTrain = y[ttindices[:ntrain], :]
xTrain = x[ttindices[:ntrain], :]
yTest = y[ttindices[ntrain+1:], :]