def map_features(votes, reviews, users, users_sim, users_conn, trusts):
""" Maps all features related to each vote, in the same order.
Args:
votes: list of vote dictionaries.
reviews: dictionary of review dictionaries indexed by review id.
users: dictionary of user dictionaries indexed by user id.
users_sim: dictionary of user similarity dictionaries indexed by a
2-tuple of user ids.
users_conn: dictionary of user connection dictionaries indexed by a
2-tuple of user ids.
trusts: networkx DiGraph with trust network.
Returns:
A dictionary of features indexed by related entity name (e.g.: voter)
and containing a list with a feature array for each vote.
"""
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(users_sim)
avg_conn = compute_avg_model(users_conn)
features = {'review': [], 'author': [], 'voter': [], 'sim': [], 'conn': []}
for vote in votes:
r_id, a_id, v_id = vote['review'], vote['author'], vote['voter']
r_feat = map_review_features(reviews[r_id])
features['review'].append(r_feat)
author = users[a_id] if a_id in users else avg_user
a_feat = map_author_features(author, avg_user)
features['author'].append(a_feat)
voter = users[v_id] if v_id in users else avg_user
v_feat = map_voter_features(voter, avg_user)
features['voter'].append(v_feat)
if v_id in users and a_id in users[v_id]['similars']:
if (a_id, v_id) in users_sim:
sim = users_sim[(a_id, v_id)]
sim_feat = map_users_sim_features(sim, avg_sim)
features['sim'].append(sim_feat)
if v_id in trusts and a_id in trusts[v_id]:
if (a_id, v_id) in users_conn:
conn = users_conn[(a_id, v_id)]
conn_feat = map_users_conn_features(conn, avg_conn)
features['conn'].append(conn_feat)
return features
def map_features(votes, reviews, users, users_sim, users_conn, trusts):
""" Maps all features related to each vote, in the same order.
Args:
votes: list of vote dictionaries.
reviews: dictionary of review dictionaries indexed by review id.
users: dictionary of user dictionaries indexed by user id.
users_sim: dictionary of user similarity dictionaries indexed by a
2-tuple of user ids.
users_conn: dictionary of user connection dictionaries indexed by a
2-tuple of user ids.
trusts: networkx DiGraph with trust network.
Returns:
A dictionary of features indexed by related entity name (e.g.: voter)
and containing a list with a feature array for each vote.
"""
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(users_sim)
avg_conn = compute_avg_model(users_conn)
features = {'review': [], 'author': [], 'voter': [], 'sim': [], 'conn': []}
for vote in votes:
r_id, a_id, v_id = vote['review'], vote['author'], vote['voter']
r_feat = map_review_features(reviews[r_id])
features['review'].append(r_feat)
author = users[a_id] if a_id in users else avg_user
a_feat = map_author_features(author, avg_user)
features['author'].append(a_feat)
voter = users[v_id] if v_id in users else avg_user
v_feat = map_voter_features(voter, avg_user)
features['voter'].append(v_feat)
if v_id in users and a_id in users[v_id]['similars']:
if (a_id, v_id) in users_sim:
sim = users_sim[(a_id, v_id)]
sim_feat = map_users_sim_features(sim, avg_sim)
features['sim'].append(sim_feat)
if v_id in trusts and a_id in trusts[v_id]:
if (a_id, v_id) in users_conn:
conn = users_conn[(a_id, v_id)]
conn_feat = map_users_conn_features(conn, avg_conn)
features['conn'].append(conn_feat)
return features
def main():
""" Main method, which performs prediction and outputs to file.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
print 'Modeling'
X_train = model_dyad(train, sim, conn, avg_sim, avg_conn)
X_val = model_dyad(val, sim, conn, avg_sim, avg_conn)
X_test = model_dyad(test, sim, conn, avg_sim, avg_conn)
train_reviews = set([v['review'] for v in train])
test_reviews = set([v['review'] for v in val
]).union(set([v['review'] for v in test]))
X_item_train, item_train_key , X_item_test, item_test_key = \
model_items(reviews, users, train_reviews, test_reviews, avg_user)
# train, test: same file, different scaling
train_users = set([v['voter'] for v in train])
test_users = set([v['voter']
for v in val]).union(set([v['voter'] for v in test]))
X_user_train, user_train_key, X_user_test, user_test_key = \
model_users(users, train_users, test_users, avg_user)
print 'Scaling'
dyad_scaler = fit_scaler('minmax', X_train)
X_train = scale_features(dyad_scaler, X_train)
X_val = scale_features(dyad_scaler, X_val)
X_test = scale_features(dyad_scaler, X_test)
item_scaler = fit_scaler('minmax', X_item_train)
X_item_train = scale_features(item_scaler, X_item_train)
X_item_test = scale_features(item_scaler, X_item_test)
user_scaler = fit_scaler('minmax', X_user_train)
X_user_train = scale_features(user_scaler, X_user_train)
X_user_test = scale_features(user_scaler, X_user_test)
X_item = vstack((X_item_train, X_item_test))
item_key = item_train_key + item_test_key
X_user = vstack((X_user_train, X_user_test))
user_key = user_train_key + user_test_key
print 'Outputting model'
output_dyad('train', train, X_train, i)
output_dyad('val', val, X_val, i)
output_dyad('test', test, X_test, i)
output_entity('item', X_item, item_key, i)
output_entity('user', X_user, user_key, i)
for j in xrange(REP):
print 'Fitting model'
print getoutput(
('Rscript lib/rlfm/rlfm_fit.R %d %d %d %d %s %d %d '
'%s') %
(_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
print getoutput(
'Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s train' %
(_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
predfile = open(
'%s/rlfm-%s-%d-%d.dat' % (_TRAIN_DIR, _CONF_STR, i, j), 'r')
pred = [float(p.strip()) for p in predfile]
predfile.close()
truth = [v['vote'] for v in train]
print len(pred)
print len(truth)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, truth)
print 'nDCG@%d: %f' % (RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred,
truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput(
'Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s val' %
(_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
print 'Predicting in test'
print getoutput(
'Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s test' %
(_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
def main():
""" Predicts votes by applying LambdaMART technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim,
avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Outputting model'
outfile = open('%s/rank_train-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
train_index = output_model(X_train, y_train, qid_train, outfile)
outfile.close()
outfile = open('%s/rank_val-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
val_index = output_model(X_val, None, qid_val, outfile)
outfile.close()
outfile = open('%s/rank_test-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
test_index = output_model(X_test, None, qid_test, outfile)
outfile.close()
for j in xrange(REP):
print 'Fitting model'
print getoutput((
'java -jar lib/ranklib/RankLib.jar -train '
'%s/rank_train-%s-%d.dat -save %s/lambdamart_model-%s-%d-%d.dat '
'-gmax 5 -ranker 6 -metric2t NDCG@5 -tree %d -leaf %d -shrinkage '
'%f') % (_DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i, j,
_T, _L, _ALPHA))
print 'Evaluating in train'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_train-%s-%d.dat '
'-score %s/rank_pred_train-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
raw_pred = []
predfile = open(
'%s/rank_pred_train-%s-%d-%d.dat' %
(_DATA_DIR, _CONF_STR, i, j), 'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in train_index]
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, j)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_val-%s-%d.dat '
'-score %s/rank_pred_val-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
predfile = open(
'%s/rank_pred_val-%s-%d-%d.dat' % (_DATA_DIR, _CONF_STR, i, j),
'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in val_index]
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open(
'%s/lambdamart-%s-%d-%d.dat' % (_VAL_DIR, _CONF_STR, i, j),
'w')
for p in pred:
print >> output, p
output.close()
print 'Predicting in test'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_test-%s-%d.dat '
'-score %s/rank_pred_test-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
predfile = open(
'%s/rank_pred_test-%s-%d-%d.dat' %
(_DATA_DIR, _CONF_STR, i, j), 'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in test_index]
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open(
'%s/lambdamart-%s-%d-%d.dat' % (_OUTPUT_DIR, _CONF_STR, i, j),
'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts votes by applying a LR regressor technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim,
avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
model = Ridge(alpha=_BETA)
# for standardized notation across algorithms, we consider alpha to be
# learning rate of and beta, regularization weight
model.fit(X_train, y_train)
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open(
'%s/lr-r:%f,f:%s,b:%s-%d-%d.dat' %
(_VAL_DIR, _BETA, _FEAT_TYPE, 'y' if _BIAS else 'n', i, 0), 'w')
for p in pred:
print >> output, p
output.close()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open(
'%s/lr-r:%f,f:%s,b:%s,-%d-%d.dat' %
(_OUTPUT_DIR, _BETA, _FEAT_TYPE, 'y' if _BIAS else 'n', i, 0), 'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts helpfulness votes using MF.
Args:
None.
Returns:
None. Results are printed to files.
"""
load_args()
for i in xrange(NUM_SETS):
t = time()
print 'Reading pickles'
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
reviews = load(open('%s/new-reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/new-sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/new-conn-%d.pkl' % (_PKL_DIR, i), 'r'))
truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val, avg_user,
avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Formatting input time: %f' % (time() - t)
for j in xrange(REP):
print 'Fitting Model'
t = time()
model = LR_Model()
model.fit(X_train, y_train, qid_train)
print 'Learning time: %f' % (time() - t)
print 'Coefficients:'
print model.w
print 'Calculating Predictions'
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print 'TRAINING ERROR'
print '-- RMSE: %f' % calculate_rmse(pred, truth)
print '-- nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews,
pred, truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
print 'Outputting validation prediction'
output = open('%s/corasvr-%s-%d-%d.dat' % (_VAL_DIR, _CONF_STR, i, j), 'w')
for p in pred:
print >> output, p
output.close()
t = time()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
print 'Prediction time: %f' % (time() - t)
print 'Outputting testing prediction'
output = open('%s/corasvr-%s-%d-%d.dat' % (_OUTPUT_DIR, _CONF_STR, i, j),
'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts votes by applying LambdaMART technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn,
test, avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Outputting model'
outfile = open('%s/rank_train-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
train_index = output_model(X_train, y_train, qid_train, outfile)
outfile.close()
outfile = open('%s/rank_val-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
val_index = output_model(X_val, None, qid_val, outfile)
outfile.close()
outfile = open('%s/rank_test-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
test_index = output_model(X_test, None, qid_test, outfile)
outfile.close()
for j in xrange(REP):
print 'Fitting model'
print getoutput(('java -jar lib/ranklib/RankLib.jar -train '
'%s/rank_train-%s-%d.dat -save %s/lambdamart_model-%s-%d-%d.dat '
'-gmax 5 -ranker 6 -metric2t NDCG@5 -tree %d -leaf %d -shrinkage '
'%f') % (_DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i, j, _T,
_L, _ALPHA))
print 'Evaluating in train'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_train-%s-%d.dat '
'-score %s/rank_pred_train-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
raw_pred = []
predfile = open('%s/rank_pred_train-%s-%d-%d.dat' % (_DATA_DIR, _CONF_STR,
i, j), 'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in train_index]
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, j)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews,
pred, train_truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_val-%s-%d.dat '
'-score %s/rank_pred_val-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
predfile = open('%s/rank_pred_val-%s-%d-%d.dat' % (_DATA_DIR, _CONF_STR,
i, j), 'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in val_index]
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/lambdamart-%s-%d-%d.dat' % (_VAL_DIR, _CONF_STR, i, j),
'w')
for p in pred:
print >> output, p
output.close()
print 'Predicting in test'
print getoutput(('java -jar lib/ranklib/RankLib.jar -load '
'%s/lambdamart_model-%s-%d-%d.dat -rank %s/rank_test-%s-%d.dat '
'-score %s/rank_pred_test-%s-%d-%d.dat -gmax 5 -metric2T NDCG@5') % \
(_MODEL_DIR, _CONF_STR, i, j, _DATA_DIR, _CONF_STR, i, _DATA_DIR,
_CONF_STR, i, j))
predfile = open('%s/rank_pred_test-%s-%d-%d.dat' % (_DATA_DIR, _CONF_STR,
i, j), 'r')
raw_pred = [float(p.strip().split()[2]) for p in predfile]
predfile.close()
pred = [raw_pred[k] for k in test_index]
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open('%s/lambdamart-%s-%d-%d.dat' % (_OUTPUT_DIR, _CONF_STR, i,
j), 'w')
for p in pred:
print >> output, p
output.close()
def predict():
""" Predicts votes by applying a SVR regressor technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
t = time()
print 'Reading data'
reviews = load(open('%s/new-reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/new-sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/new-conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val, avg_user,
avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Formatting input time: %f' % (time() - t)
t = time()
model = SVR(C=_C, epsilon=_EPS, kernel=_KERNEL)
model.fit(X_train , y_train)
print 'Learning time: %f' % (time() - t)
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/svr-c:%f,k:%s,e:%f,f:%s,b:%s-%d-%d.dat' % (_VAL_DIR,
_C, _KERNEL, _EPS, _FEAT_TYPE, 'y' if _BIAS else 'n', i, 0),
'w')
for p in pred:
print >> output, p
output.close()
t = time()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
print 'Prediction time: %f' % (time() - t)
output = open('%s/svr-c:%f,k:%s,e:%f,f:%s,b:%s-%d-%d.dat' %
(_OUTPUT_DIR, _C, _KERNEL, _EPS, _FEAT_TYPE,
'y' if _BIAS else 'n', i, 0), 'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts votes by applying a LR regressor technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn,
test, avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
model = Ridge(alpha=_BETA)
# for standardized notation across algorithms, we consider alpha to be
# learning rate of and beta, regularization weight
model.fit(X_train , y_train)
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/lr-r:%f,f:%s,b:%s-%d-%d.dat' % (_VAL_DIR, _BETA,
_FEAT_TYPE, 'y' if _BIAS else 'n', i, 0), 'w')
for p in pred:
print >> output, p
output.close()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open('%s/lr-r:%f,f:%s,b:%s,-%d-%d.dat' % (_OUTPUT_DIR, _BETA,
_FEAT_TYPE, 'y' if _BIAS else 'n', i, 0), 'w')
for p in pred:
print >> output, p
output.close()
def predict():
""" Predicts votes by applying RankSVM technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn,
test, avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Outputting model'
outfile = open('%s/rank_train-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
train_index = output_model(X_train, y_train, qid_train, outfile)
outfile.close()
outfile = open('%s/rank_val-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
val_index = output_model(X_val, None, qid_val, outfile)
outfile.close()
outfile = open('%s/rank_test-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i), 'w')
test_index = output_model(X_test, None, qid_test, outfile)
outfile.close()
print 'Fitting model'
print getoutput(('lib/svm_rank/svm_rank_learn -c %f -w %s -t %s '
'%s/rank_train-%s-%d.dat %s/rank_model-%s-%d-0.dat') % (_C, _ALGO,
_KERNEL, _DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i))
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_train-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_train-%s-%d-0.dat') % (_DATA_DIR,
_CONF_STR, i, _MODEL_DIR, _CONF_STR, i, _DATA_DIR, _CONF_STR, i))
raw_pred = []
predfile = open('%s/rank_pred_train-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR,
i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in train_index]
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews,
pred, train_truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_val-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_val-%s-%d-0.dat') % (_DATA_DIR,
_CONF_STR, i, _MODEL_DIR, _CONF_STR, i, _DATA_DIR, _CONF_STR, i))
predfile = open('%s/rank_pred_val-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR,
i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in val_index]
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/svmrank-%s-%d-0.dat' % (_VAL_DIR, _CONF_STR, i), 'w')
for p in pred:
print >> output, p
output.close()
print 'Predicting in test'
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_test-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_test-%s-%d-0.dat') % (_DATA_DIR,
_CONF_STR, i, _MODEL_DIR, _CONF_STR, i, _DATA_DIR, _CONF_STR, i))
predfile = open('%s/rank_pred_test-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR,
i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in test_index]
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open('%s/svmrank-%s-%d-0.dat' % (_OUTPUT_DIR, _CONF_STR, i), 'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts helpfulness votes using MF.
Args:
None.
Returns:
None. Results are printed to files.
"""
load_args()
for i in xrange(NUM_SETS):
t = time()
print 'Reading pickles'
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
reviews = load(open('%s/new-reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/new-sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/new-conn-%d.pkl' % (_PKL_DIR, i), 'r'))
truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim,
avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Formatting input time: %f' % (time() - t)
for j in xrange(REP):
print 'Fitting Model'
t = time()
model = LR_Model()
model.fit(X_train, y_train, qid_train)
print 'Learning time: %f' % (time() - t)
print 'Coefficients:'
print model.w
print 'Calculating Predictions'
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print 'TRAINING ERROR'
print '-- RMSE: %f' % calculate_rmse(pred, truth)
print '-- nDCG@%d: %f' % (
RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred, truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
print 'Outputting validation prediction'
output = open(
'%s/corasvr-%s-%d-%d.dat' % (_VAL_DIR, _CONF_STR, i, j), 'w')
for p in pred:
print >> output, p
output.close()
t = time()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
print 'Prediction time: %f' % (time() - t)
print 'Outputting testing prediction'
output = open(
'%s/corasvr-%s-%d-%d.dat' % (_OUTPUT_DIR, _CONF_STR, i, j),
'w')
for p in pred:
print >> output, p
output.close()
def predict():
""" Predicts votes by applying RankSVM technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim,
avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Outputting model'
outfile = open('%s/rank_train-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
train_index = output_model(X_train, y_train, qid_train, outfile)
outfile.close()
outfile = open('%s/rank_val-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
val_index = output_model(X_val, None, qid_val, outfile)
outfile.close()
outfile = open('%s/rank_test-%s-%d.dat' % (_DATA_DIR, _CONF_STR, i),
'w')
test_index = output_model(X_test, None, qid_test, outfile)
outfile.close()
print 'Fitting model'
print getoutput(('lib/svm_rank/svm_rank_learn -c %f -w %s -t %s '
'%s/rank_train-%s-%d.dat %s/rank_model-%s-%d-0.dat') %
(_C, _ALGO, _KERNEL, _DATA_DIR, _CONF_STR, i,
_MODEL_DIR, _CONF_STR, i))
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_train-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_train-%s-%d-0.dat') %
(_DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i,
_DATA_DIR, _CONF_STR, i))
raw_pred = []
predfile = open(
'%s/rank_pred_train-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR, i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in train_index]
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_val-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_val-%s-%d-0.dat') %
(_DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i,
_DATA_DIR, _CONF_STR, i))
predfile = open(
'%s/rank_pred_val-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR, i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in val_index]
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/svmrank-%s-%d-0.dat' % (_VAL_DIR, _CONF_STR, i), 'w')
for p in pred:
print >> output, p
output.close()
print 'Predicting in test'
print getoutput(('lib/svm_rank/svm_rank_classify '
'%s/rank_test-%s-%d.dat %s/rank_model-%s-%d-0.dat '
'%s/rank_pred_test-%s-%d-0.dat') %
(_DATA_DIR, _CONF_STR, i, _MODEL_DIR, _CONF_STR, i,
_DATA_DIR, _CONF_STR, i))
predfile = open(
'%s/rank_pred_test-%s-%d-0.dat' % (_DATA_DIR, _CONF_STR, i), 'r')
raw_pred = [float(p.strip()) for p in predfile]
predfile.close()
pred = [raw_pred[j] for j in test_index]
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open('%s/svmrank-%s-%d-0.dat' % (_OUTPUT_DIR, _CONF_STR, i),
'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Main method, which performs prediction and outputs to file.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
print 'Creating average user (for mean imputation)'
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
print 'Modeling'
X_train = model_dyad(train, sim, conn, avg_sim, avg_conn)
X_val = model_dyad(val, sim, conn, avg_sim, avg_conn)
X_test = model_dyad(test, sim, conn, avg_sim, avg_conn)
train_reviews = set([v['review'] for v in train])
test_reviews = set([v['review'] for v in val]).union(set([v['review'] for v
in test]))
X_item_train, item_train_key , X_item_test, item_test_key = \
model_items(reviews, users, train_reviews, test_reviews, avg_user)
# train, test: same file, different scaling
train_users = set([v['voter'] for v in train])
test_users = set([v['voter'] for v in val]).union(set([v['voter'] for v in
test]))
X_user_train, user_train_key, X_user_test, user_test_key = \
model_users(users, train_users, test_users, avg_user)
print 'Scaling'
dyad_scaler = fit_scaler('minmax', X_train)
X_train = scale_features(dyad_scaler, X_train)
X_val = scale_features(dyad_scaler, X_val)
X_test = scale_features(dyad_scaler, X_test)
item_scaler = fit_scaler('minmax', X_item_train)
X_item_train = scale_features(item_scaler, X_item_train)
X_item_test = scale_features(item_scaler, X_item_test)
user_scaler = fit_scaler('minmax', X_user_train)
X_user_train = scale_features(user_scaler, X_user_train)
X_user_test = scale_features(user_scaler, X_user_test)
X_item = vstack((X_item_train, X_item_test))
item_key = item_train_key + item_test_key
X_user = vstack((X_user_train, X_user_test))
user_key = user_train_key + user_test_key
print 'Outputting model'
output_dyad('train', train, X_train, i)
output_dyad('val', val, X_val, i)
output_dyad('test', test, X_test, i)
output_entity('item', X_item, item_key, i)
output_entity('user', X_user, user_key, i)
for j in xrange(REP):
print 'Fitting model'
print getoutput(('Rscript lib/rlfm/rlfm_fit.R %d %d %d %d %s %d %d '
'%s') % (_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
print getoutput('Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s train' % (_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
predfile = open('%s/rlfm-%s-%d-%d.dat' % (_TRAIN_DIR, _CONF_STR, i,
j), 'r')
pred = [float(p.strip()) for p in predfile]
predfile.close()
truth = [v['vote'] for v in train]
print len(pred)
print len(truth)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews, pred,
truth, RANK_SIZE))
print 'Predicting in validation'
print getoutput('Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s val' % (_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
print 'Predicting in test'
print getoutput('Rscript lib/rlfm/rlfm_predict.R %d %d %d %d %s %d %d '
'%s test' % (_K, _ITER, _SAMPLES, _BURN_IN, _FEAT, i, j, _DATA_DIR))
def predict():
""" Predicts votes by applying a SVR regressor technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
t = time()
print 'Reading data'
reviews = load(open('%s/new-reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/new-sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/new-conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim,
avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val,
avg_user, avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
print 'Formatting input time: %f' % (time() - t)
t = time()
model = SVR(C=_C, epsilon=_EPS, kernel=_KERNEL)
model.fit(X_train, y_train)
print 'Learning time: %f' % (time() - t)
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, 0)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE,
calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open(
'%s/svr-c:%f,k:%s,e:%f,f:%s,b:%s-%d-%d.dat' %
(_VAL_DIR, _C, _KERNEL, _EPS, _FEAT_TYPE, 'y' if _BIAS else 'n', i,
0), 'w')
for p in pred:
print >> output, p
output.close()
t = time()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
print 'Prediction time: %f' % (time() - t)
output = open(
'%s/svr-c:%f,k:%s,e:%f,f:%s,b:%s-%d-%d.dat' %
(_OUTPUT_DIR, _C, _KERNEL, _EPS, _FEAT_TYPE, 'y' if _BIAS else 'n',
i, 0), 'w')
for p in pred:
print >> output, p
output.close()
def main():
""" Predicts votes by applying a GBRT regressor technique.
Args:
None.
Returns:
None.
"""
load_args()
for i in xrange(NUM_SETS):
print 'Reading data'
reviews = load(open('%s/reviews-%d.pkl' % (_PKL_DIR, i), 'r'))
users = load(open('%s/users-%d.pkl' % (_PKL_DIR, i), 'r'))
train = load(open('%s/train-%d.pkl' % (_PKL_DIR, i), 'r'))
test = load(open('%s/test-%d.pkl' % (_PKL_DIR, i), 'r'))
val = load(open('%s/validation-%d.pkl' % (_PKL_DIR, i), 'r'))
sim = load(open('%s/sim-%d.pkl' % (_PKL_DIR, i), 'r'))
conn = load(open('%s/conn-%d.pkl' % (_PKL_DIR, i), 'r'))
train_truth = [v['vote'] for v in train]
if _BIAS:
bias = BiasModel()
train = bias.fit_transform(train, reviews)
avg_user = compute_avg_user(users)
avg_sim = compute_avg_model(sim)
avg_conn = compute_avg_model(conn)
X_train, y_train, qid_train = generate_input(reviews, users, sim, conn,
train, avg_user, avg_sim, avg_conn)
X_val, _, qid_val = generate_input(reviews, users, sim, conn, val, avg_user,
avg_sim, avg_conn)
X_test, _, qid_test = generate_input(reviews, users, sim, conn, test,
avg_user, avg_sim, avg_conn)
scaler = fit_scaler('minmax', X_train)
X_train = scale_features(scaler, X_train)
X_val = scale_features(scaler, X_val)
X_test = scale_features(scaler, X_test)
for j in xrange(REP):
model = GradientBoostingRegressor(loss=_LOSS, learning_rate=_ALPHA,
n_estimators=_T, max_depth=_MAX_D, subsample=_SUBSAMPLE,
max_features=_MAX_F, random_state=(int(time() * 1000000) % 1000000))
model.fit(X_train, y_train)
pred = model.predict(X_train)
if _BIAS:
bias.add_bias(train, reviews, pred)
print '~ Training error on set %d repetition %d' % (i, j)
print 'RMSE: %f' % calculate_rmse(pred, train_truth)
print 'nDCG@%d: %f' % (RANK_SIZE, calculate_avg_ndcg(train, reviews, pred,
train_truth, RANK_SIZE))
pred = model.predict(X_val)
if _BIAS:
bias.add_bias(val, reviews, pred)
output = open('%s/gbrt-l:%f,t:%d,d:%d,e:%s,p:%f,m:%s,f:%s,b:%s-%d-%d.dat' %
(_VAL_DIR, _ALPHA, _T, _MAX_D, _LOSS, _SUBSAMPLE, str(_MAX_F),
_FEAT_TYPE, 'y' if _BIAS else 'n', i, j), 'w')
for p in pred:
print >> output, p
output.close()
pred = model.predict(X_test)
if _BIAS:
bias.add_bias(test, reviews, pred)
output = open('%s/gbrt-l:%f,t:%d,d:%d,e:%s,p:%f,m:%s,f:%s,b:%s-%d-%d.dat' %
(_OUTPUT_DIR, _ALPHA, _T, _MAX_D, _LOSS, _SUBSAMPLE, str(_MAX_F),
_FEAT_TYPE, 'y' if _BIAS else 'n', i, j), 'w')
for p in pred:
print >> output, p
output.close()