def pyFM_cv(verbose=True, t=Timer()):
#pyFM parameters
factors = np.linspace(20, 200, 10, dtype=np.int64)
learning_rates = np.logspace(-3, -3, 1)
params = dict()
rmses = dict()
for k in factors:
params['k'] = k
for rate in learning_rates:
params['rate'] = rate
algo = pylibfm.FM(num_factors=k,
num_iter=200,
verbose=True,
task="regression",
initial_learning_rate=rate,
learning_rate_schedule="optimal")
rmse = pyFM_cv_algo(algo)
print(
"------Time:{}, rmse: {}, factors: {}, learning_rates: {}------\n\n"
.format(t.now(), rmse, k, rate))
rmses[rmse] = params
# Find the model with least RMSE
lowest_rmse = min(rmses.keys())
best_params = rmses[lowest_rmse]
print("Best pyFM rmse: {}. Params: factors: {}, learning_rates: {}".format(
lowest_rmse, best_params['k'], best_params['rate']))
def pyFMJob(data_path,
params,
N,
vectorizer,
with_timestamps=False,
with_authors=False):
rmses = []
logging.info("Evaluando con params: {0}".format(params))
for i in range(1, 4 + 1):
train_data, y_tr, _ = loadData('train/train_N' + str(N) + '.' + str(i),
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
X_tr = vectorizer.transform(train_data)
fm = pylibfm.FM(num_factors=params['f'], num_iter=params['mi'], k0=params['bias'], k1=params['oneway'], init_stdev=params['init_stdev'], \
validation_size=params['val_size'], learning_rate_schedule=params['lr_s'], initial_learning_rate=params['lr'], \
power_t=params['invscale_pow'], t0=params['optimal_denom'], shuffle_training=params['shuffle'], seed=params['seed'], \
task='regression', verbose=True)
fm.fit(X_tr, y_tr)
val_data, y_va, _ = loadData('val/val_N' + str(N) + '.' + str(i),
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
X_va = vectorizer.transform(val_data)
preds = fm.predict(X_va)
rmse = sqrt(mean_squared_error(y_va, preds))
print("FM RMSE: %.4f" % rmse)
rmses.append(rmse)
return mean(rmses)
def pyfm_predict(train_actual, predict):
"""
Matrix Factorization using SGD with pyFM library.
Compute the predictions on a test_set after training on a train_set using the method Svd++ Matrix Factorization using SGD with pyFM library
Args:
train_actual (pandas.DataFrame): train set
predict (pandas.DataFrame): test set
Hyperparameters:
num_factors : The number of factors.
num_iter : The number of iteration of the SGD procedure
initial_learning_rate:
Returns:
numpy array: predictions
"""
print("pyfm")
predict_data, y_predict = create_input_pyfm(predict)
train_actual_data, y_train_actual = create_input_pyfm(train_actual)
v = DictVectorizer()
X_train = v.fit_transform(train_actual_data)
X_test = v.transform(predict_data)
# Hyperparameters
num_factors = 20
num_iter = 200
task = 'regression'
initial_learning_rate = 0.001
learning_rate_schedule = 'optimal'
fm = pylibfm.FM(num_factors=num_factors,
num_iter=num_iter,
task=task,
initial_learning_rate=initial_learning_rate,
learning_rate_schedule=learning_rate_schedule)
fm.fit(X_train, y_train_actual)
preds = fm.predict(X_test)
return np.clip(preds, 1, 5)
def main(para_dname):
data_dir = 'data'
output_dir = 'output'
dataset_name = para_dname
# dataset_name = 'ml-1m'
# dataset_name = 'douban'
# dataset_name = 'yahoo_music'
train_X, test_X, train_y, test_y = load_dataset(data_dir, dataset_name)
v = DictVectorizer()
train_X = v.fit_transform(train_X)
test_X = v.transform(test_X)
fm = pylibfm.FM(num_factors=10,
num_iter=1000,
verbose=True,
task='regression',
initial_learning_rate=0.001,
learning_rate_schedule='optimal',
validation_size=0.1)
train_loss, val_loss = fm.fit(train_X, train_y)
train_loss = np.sqrt(np.array(train_loss))
val_loss = np.sqrt(np.array(val_loss))
np.save(os.path.join(output_dir, dataset_name + '_trloss'), train_loss)
np.save(os.path.join(output_dir, dataset_name + '_valloss'), val_loss)
preds = fm.predict(test_X)
test_loss = math.sqrt(mean_squared_error(test_y, preds))
print(preds)
print('Test loss: %.5f' % test_loss)
return 0
def __init__(self,
iter=100,
factor=10,
use_info=True,
path='./',
external_fm=None):
from pyfm import pylibfm
self.__use_info = use_info
# temp code, load ml-100k's info
if self.__use_info:
self.__info = Info(path)
# Build and train a Factorization Machine
if external_fm:
print >> sys.stderr, 'Use external FM: %s' % type(external_fm)
self.__fm = external_fm
else:
print >> sys.stderr, 'iter=%d, factor=%d, use_info=%d' % (
iter, factor, use_info)
self.__fm = pylibfm.FM(num_factors=factor,
num_iter=iter,
verbose=True,
task="regression",
initial_learning_rate=0.001,
learning_rate_schedule="optimal")
def FM(train_data, test_data, y_train):
import numpy as np
from sklearn.feature_extraction import DictVectorizer
from pyfm import pylibfm
import pandas as pd
import math
import time
start_time = time.time()
# Transforms lists of feature-value mappings to one-hot encoded vectors
v = DictVectorizer()
X_train = v.fit_transform(train_data)
X_test = v.transform(test_data)
# Build and train a Factorization Machine
fm = pylibfm.FM(num_factors=5,
num_iter=10,
verbose=True,
task="regression",
initial_learning_rate=0.1,
learning_rate_schedule="optimal")
fm.fit(X_train, y_train)
preds = fm.predict(X_test)
# Time taken for training (seconds)
print("--- %s seconds : Time taken for training ---" %
(time.time() - start_time))
np.savetxt('predictions.txt', preds)
return preds
def __init__(self, n_factor=16, n_iter=10, use_attrs=True):
self.fm = pylibfm.FM(num_factors=n_factor,
num_iter=n_iter,
task="regression",
initial_learning_rate=0.001,
learning_rate_schedule="optimal")
self.use_attrs = use_attrs
self.v = DictVectorizer()
def cf_models(cf_lib, N, data_path, params):
#caching, ya que son siempre los mismos params del CF
cf_models = {}
all_data, y_all, items = loadData(
"eval_all_N" + str(N) + ".data",
data_path='TwitterRatings/funkSVD/data_with_authors/',
with_timestamps=False,
with_authors=True) #hardcoded bc f**k it
if cf_lib == "pyFM":
v = DictVectorizer()
X_all = v.fit_transform(all_data)
for i in range(1, 4 + 1):
train_data, y_tr, _ = loadData(
'train/train_N' + str(N) + '.' + str(i),
data_path='TwitterRatings/funkSVD/data_with_authors/',
with_timestamps=False,
with_authors=True)
X_tr = v.transform(train_data)
fm = pylibfm.FM(num_factors=params['f'], num_iter=params['mi'], k0=params['bias'], k1=params['oneway'], init_stdev=params['init_stdev'], \
validation_size=params['val_size'], learning_rate_schedule=params['lr_s'], initial_learning_rate=params['lr'], \
power_t=params['invscale_pow'], t0=params['optimal_denom'], shuffle_training=params['shuffle'], seed=params['seed'], \
task='regression', verbose=True)
fm.fit(X_tr, y_tr)
cf_models[i] = fm
return cf_models, v, items
elif cf_lib == "implicit":
all_c = consumption(ratings_path=data_path + 'eval_all_N' + str(N) +
'.data',
rel_thresh=0,
with_ratings=True)
items_ids = list(
set([
itemId for userId, itemsDict in all_c.items()
for itemId in itemsDict
]))
idcoder = IdCoder(items_ids, all_c.keys())
for i in range(1, 4 + 1):
ones, row, col = get_data(data_path=data_path,
all_c=all_c,
idcoder=idcoder,
fold=i,
N=N,
mode="tuning")
matrix = csr_matrix((ones, (row, col)), dtype=np.float64)
user_items = matrix.T.tocsr()
model = implicit.als.AlternatingLeastSquares(
factors=params['f'],
regularization=params['lamb'],
iterations=params['mi'],
dtype=np.float64)
model.fit(matrix)
cf_models[i] = model
return cf_models, idcoder, items
def train(self, n_epochs: int, learning_rate: float = 0.001,
random_seed: int = 42, hybrid: bool = False, verbose: bool = True):
self._build_train_test_ds(hybrid)
self.fm = pylibfm.FM(num_factors=self.k,
num_iter=n_epochs,
verbose=verbose,
task='regression',
initial_learning_rate=learning_rate,
seed=random_seed)
self.fm.fit(self.X_train, self.y_train)
def __init__(self):
self.fm = pylibfm.FM(k1=False,
validation_size=0.005,
num_factors=10,
num_iter=8,
verbose=True,
task="classification",
initial_learning_rate=0.001,
init_stdev=0.002,
learning_rate_schedule="optimal")
def train(training_data, labels):
# Use one hot-encoding
label_encoder = LabelEncoder()
vectorizer = DictVectorizer()
train_event_x = vectorizer.fit_transform(training_data)
train_event_y = label_encoder.fit_transform(labels)
# Create and train the model.
pctr_estimator = pylibfm.FM()
pctr_estimator.fit(train_event_x, train_event_y)
model = (pctr_estimator, label_encoder, vectorizer)
print('Training done')
return model
def __init__(self, rec_name, dataset, uses_features):
super(FMRec, self).__init__(rec_name, dataset, uses_features)
# init
self.one_hot_columns = None
# default rec
self.fm = pylibfm.FM(num_factors=50,
num_iter=10,
task="regression",
initial_learning_rate=0.001,
learning_rate_schedule="optimal",
verbose=True)
def train(training_data, labels):
#label encoder features to get prediction
label_encoder = LabelEncoder()
vectorizer = DictVectorizer()
train_event_x = vectorizer.fit_transform(training_data)
train_event_y = label_encoder.fit_transform(labels)
#definition of balanced weights
global weights
weights = len(labels) / (sum(labels) * 2)
# Create and train the model using Factorization Machine Algorithm
pctr_estimator = pylibfm.FM()
pctr_estimator.fit(train_event_x, train_event_y)
model = (pctr_estimator, label_encoder, vectorizer)
print('Training done')
return model
def __init__(self,
num_factors=10,
num_iter=1,
k0=True,
k1=True,
init_stdev=0.1,
validation_size=0.01,
learning_rate_schedule="optimal",
initial_learning_rate=0.01,
power_t=0.5,
t0=0.001,
task='classification',
verbose=True,
shuffle_training=True,
seed=28):
super(BaseEstimator, self).__init__()
self.num_factors = num_factors
self.num_iter = num_iter
self.k0 = k0
self.k1 = k1
self.init_stdev = init_stdev
self.validation_size = validation_size
self.learning_rate_schedule = learning_rate_schedule
self.initial_learning_rate = initial_learning_rate
self.power_t = power_t
self.t0 = t0
self.task = task
self.verbose = verbose
self.shuffle_training = shuffle_training
self.seed = seed
self.fm = pylibfm.FM(
num_factors=self.num_factors,
num_iter=self.num_iter,
k0=self.k0,
k1=self.k1,
init_stdev=self.init_stdev,
validation_size=self.validation_size,
learning_rate_schedule=self.learning_rate_schedule,
initial_learning_rate=self.initial_learning_rate,
power_t=self.power_t,
t0=self.t0,
task=self.task,
verbose=self.verbose,
shuffle_training=self.shuffle_training,
seed=self.seed)
def main():
print(getCurrentTime(), "running...")
dok_train = np.load(r"%s\..\input\sparse_train.npy" % runningPath)[()]# train 中包括 '2018-09-17'-- '2018-09-23' 的数据
dok_verify = np.load(r"%s\..\input\sparse_verify.npy" % runningPath)[()] # verify 中只包含了 '2018-09-24' 的数据
dok_test = np.load(r"%s\..\input\sparse_test.npy" % runningPath)[()]
train_label = pd.read_csv(r'%s\..\input\train_label.txt' % runningPath)
verify_label = pd.read_csv(r'%s\..\input\verify_label.txt' % runningPath)
fm = pylibfm.FM(num_factors=50, num_iter=10, verbose=True, task="classification", initial_learning_rate=0.0001, learning_rate_schedule="optimal")
fm.fit(dok_train, train_label['is_trade'])
Y_prediced = fm.predict(dok_verify)
pyfm_logloss = -np.sum(verify_label * np.log(Y_prediced) +
(1 - verify_label) * np.log(1 - Y_prediced))/ Y_prediced.shape[0]
print(getCurrentTime(), "lgb logloss %.6f" %(pyfm_logloss))
return
def pyfm(train, test, **arg):
print('[PYFM] applying')
# Get the args
num_factors = arg['num_factors']
num_iter = arg['num_iter']
task = arg['task']
initial_learning_rate = arg['initial_learning_rate']
learning_rate_schedule = arg['learning_rate_schedule']
(train_data, y_train, train_users, train_items) = prepare_data(train)
(test_data, y_test, test_users, test_items) = prepare_data(test)
v = DictVectorizer()
X_train = v.fit_transform(train_data)
X_test = v.transform(test_data)
fm = pylibfm.FM(num_factors=num_factors,
num_iter=num_iter,
task=task,
initial_learning_rate=initial_learning_rate,
learning_rate_schedule=learning_rate_schedule)
fm.fit(X_train, y_train)
preds = fm.predict(X_test)
for i in range(len(preds)):
if preds[i] > 5:
preds[i] = 5
elif preds[i] < 1:
preds[i] = 1
df_return = test.copy()
df_return.Rating = preds
print('[PYFM] done')
return df_return
def peretrain():
X_train = []
y = []
for user in os.listdir("userinfo"):
for file in os.listdir("userinfo/" + user):
if file == "dislike.txt":
dann = [
int(i) for i in open("userinfo/" + user + '/' +
file, "r").read().split()
if checkint(i)
]
for d in dann:
X_train.append({
"who_marked": str(user),
"marked_user": str(d)
})
y.append(0)
elif file == "like.txt":
dann = [
int(i) for i in open("userinfo/" + user + '/' +
file, "r").read().split()
if checkint(i)
]
for d in dann:
X_train.append({
"who_marked": str(user),
"marked_user": str(d)
})
y.append(1)
v = DictVectorizer()
X = v.fit_transform(X_train)
y = np.array(y, dtype=np.float64)
fm = pylibfm.FM(num_factors=20,
num_iter=len(y) * 3,
verbose=False,
task="regression",
initial_learning_rate=0.001,
learning_rate_schedule="optimal")
fm.fit(X, y)
return fm
def dofit_pyfm():
d = get_subsample()
globals().update(d)
clf = pylibfm.FM(num_factors=4,
num_iter=100,
verbose=True,
task="classification",
initial_learning_rate=0.00001,
learning_rate_schedule="optimal")
scaler = preproc.StandardScaler(with_mean=False) # hmm
scaler.fit(X_train)
def transx(x):
x = scaler.transform(x)
return scipy.sparse.csr_matrix(x)
clf._fit_old = clf.fit
clf.fit = lambda x, y: clf._fit_old(transx(x), y)
clf._predict_old = clf.predict
clf.predict = lambda x: clf._predict_old(transx(x))
clf.fit(X_train, y_train)
return {'clf': clf}
def benchmark(task='regression', content=False):
losses = []
total_time = 0
for k in range(5):
print('== Fold %d' % (k+1))
# set RNG
np.random.seed(0)
random.seed(0)
# Load data
(train_data, y_train) = load_ml_100k("u%d.base" % (k+1), content)
(test_data, y_test) = load_ml_100k("u%d.test" % (k+1), content)
if task == "classification":
y_test = np.greater(y_test, 3)
# Transform to matrix
v = DictVectorizer()
x_train = v.fit_transform(train_data)
x_test = v.transform(test_data)
# Build and train a Factorization Machine
fm = pylibfm.FM(num_iter=20, verbose=True, task=task,
initial_learning_rate=0.005,
learning_rate_schedule="constant", seed=0)
start = time.time()
fm.fit(x_train, y_train)
used = time.time() - start
total_time += used
# Evaluate
predictions = fm.predict(x_test)
if task == "regression":
losses.append(root_mean_squared_error(y_test, predictions))
print("FM RMSE: %.4f" % losses[-1])
elif task == "classification":
losses.append(log_loss(y_test, predictions))
print("FM log loss: %.4f" % losses[-1])
print("Time used: %.4fs" % used)
print('== Summary')
print('Mean RMSE: %.4f' % np.mean(losses))
print('Total time: %.4fs' % total_time)
val_w = sample_weight[test]
# reg = LogisticRegression(C=0.1, solver='sag', n_jobs=-1)
# pred_x = cross_val_predict(reg, trn_x, trn_y, cv=5, n_jobs=-1)
# trn_x = np.c_[trn_x, pred_x]
"""
clf = TFFMClassifier(order=6,
rank=10,
optimizer=tf.train.AdagradOptimizer(0.01),
n_epochs=100,
batch_size=10000,
init_std=0.001,
input_type='sparse'
)
"""
clf = pylibfm.FM(**params)
clf.fit(trn_x, trn_y)
_score = log_loss(val_y, clf.predict(val_x), sample_weight=val_w)
_score2 = -roc_auc_score(
val_y, clf.predict(val_x), sample_weight=val_w)
# logger.debug(' _score: %s' % _score)
list_score.append(_score)
list_score2.append(_score2)
break
score = (np.mean(list_score), np.min(list_score), np.max(list_score))
score2 = (np.mean(list_score2), np.min(list_score2),
np.max(list_score2))
logger.info('param: %s' % (params))
v = DictVectorizer()
# X_origin = train.loc[:,['iid','uid']].astype(np.string_).to_dict(orient='records')
# X = v.fit_transform(X_origin)
y = np.array(train.loc[:, ['score']]).flatten().astype(np.float64)
#X_train, X_test, y_train, y_test = train_test_split(X, y)
X_merge = pd.concat([train.loc[:, ['uid', 'iid']], test])
X_merge_hot = v.fit_transform(
X_merge.astype(np.string_).to_dict(orient='records'))
train_hot = X_merge_hot[0:train.shape[0]]
test_hot = X_merge_hot[train.shape[0]:X_merge_hot.shape[0]]
print "data is ready"
fm = pylibfm.FM(num_factors=100,
num_iter=30,
verbose=True,
task="regression",
initial_learning_rate=0.01,
learning_rate_schedule="optimal")
#y_train = y_train.astype(np.float64)
fm.fit(train_hot, y)
print 'fit well'
joblib.dump(fm, "fm_model10,0000_iter_30.m")
print "start predict"
y_pred = fm.predict(test_hot)
df_fm = pd.DataFrame(y_pred, columns=['score'])
df_fm.to_csv("fm_result100_0000_iter_30.csv", index=False)
#
# ground_truth = np.around(preds)
from pyfm import pylibfm
from sklearn.feature_extraction import DictVectorizer
iris_data = load_iris()
X = iris_data['data']
y = iris_data['target'] == 2
data = [{v: k for k, v in dict(zip(i, range(len(i)))).items()} for i in X]
X_train, X_test, y_train, y_test = train_test_split(data,
y,
test_size=0.3,
random_state=0)
v = DictVectorizer()
X_train = v.fit_transform(X_train)
X_test = v.transform(X_test)
fm = pylibfm.FM(num_factors=50,
num_iter=1000,
verbose=True,
task="classification",
initial_learning_rate=0.0001,
learning_rate_schedule="optimal")
fm.fit(X_train, y_train)
y_preds = fm.predict(X_test)
from sklearn.metrics import log_loss
print("Validation log loss: %.4f" % log_loss(y_test, y_preds))
def hybrid_protocol_evaluation(data_path, data_path_context, cf_lib, solr,
params_cb, params_cf, params_hy, N):
test_c = consumption(ratings_path=data_path + 'test/test_N20.data',
rel_thresh=0,
with_ratings=True)
train_c = consumption(ratings_path=data_path + 'eval_train_N20.data',
rel_thresh=0,
with_ratings=False)
all_c = consumption(ratings_path=data_path + 'eval_all_N20.data',
rel_thresh=0,
with_ratings=True)
MRRs = dict((N, []) for N in [5, 10, 15, 20])
nDCGs = dict((N, []) for N in [5, 10, 15, 20])
APs = dict((N, []) for N in [5, 10, 15, 20])
Rprecs = dict((N, []) for N in [5, 10, 15, 20])
if cf_lib == "pyFM":
all_data, y_all, items = loadData("eval_all_N20.data",
data_path=data_path_context,
with_timestamps=False,
with_authors=True)
v = DictVectorizer()
X_all = v.fit_transform(all_data)
train_data, y_tr, _ = loadData('eval_train_N20.data',
data_path=data_path_context,
with_timestamps=False,
with_authors=True)
X_tr = v.transform(train_data)
fm = pylibfm.FM(num_factors=params_cf['f'], num_iter=params_cf['mi'], k0=params_cf['bias'], k1=params_cf['oneway'], init_stdev=params_cf['init_stdev'], \
validation_size=params_cf['val_size'], learning_rate_schedule=params_cf['lr_s'], initial_learning_rate=params_cf['lr'], \
power_t=params_cf['invscale_pow'], t0=params_cf['optimal_denom'], shuffle_training=params_cf['shuffle'], seed=params_cf['seed'], \
task='regression', verbose=True)
fm.fit(X_tr, y_tr)
elif cf_lib == "implicit":
items_ids = list(
set([
itemId for userId, itemsDict in all_c.items()
for itemId in itemsDict
]))
idcoder = IdCoder(items_ids, all_c.keys())
ones, row, col = get_data(data_path=data_path,
all_c=all_c,
idcoder=idcoder,
fold=0,
N=20,
mode="testing")
matrix = csr_matrix((ones, (row, col)), dtype=np.float64)
user_items = matrix.T.tocsr()
model = implicit.als.AlternatingLeastSquares(
factors=params_cf['f'],
regularization=params_cf['lamb'],
iterations=params_cf['mi'],
dtype=np.float64)
model.fit(matrix)
p = 0
for userId in test_c:
logging.info("#u: {0}/{1}".format(p, len(test_c)))
p += 1
if cf_lib == "pyFM":
user_rows = [{
'user_id': str(userId),
'item_id': str(itemId)
} for itemId in items]
X_te = v.transform(user_rows)
preds = fm.predict(X_te)
recs_cf = [
itemId for _, itemId in sorted(zip(preds, items), reverse=True)
]
elif cf_lib == "implicit":
recommends = model.recommend(userid=int(
idcoder.coder('user', userId)),
user_items=user_items,
N=200)
recs_cf = [idcoder.decoder('item', tupl[0]) for tupl in recommends]
recs_cb = []
for itemId in train_c[userId]:
encoded_params = urlencode(params_cb)
url = solr + '/mlt?q=goodreadsId:' + itemId + "&" + encoded_params
response = json.loads(urlopen(url).read().decode('utf8'))
try:
docs = response['response']['docs']
except TypeError as e:
continue
recs_cb.append([str(doc['goodreadsId'][0]) for doc in docs])
recs_cb = flatten_list(list_of_lists=recs_cb, rows=params_cb['rows'])
recs_cf = remove_consumed(user_consumption=train_c[userId],
rec_list=recs_cf)
recs_cf = recs_cf[:200]
recs_cb = remove_consumed(user_consumption=train_c[userId],
rec_list=recs_cb)
recs_cb = recs_cb[:200]
recs_hy = hybridize_recs(recs_cb=recs_cb,
recs_cf=recs_cf,
weight_cb=params_hy['weight_cb'],
weight_cf=params_hy['weight_cf'])
recs_hy = remove_consumed(user_consumption=train_c[userId],
rec_list=recs_hy)
recs_hy = recs_cleaner(solr=solr,
consumpt=train_c[userId],
recs=recs_hy[:100])
recs_hy = user_ranked_recs(user_recs=recs_hy,
user_consumpt=test_c[userId])
for N in [5, 10, 15, 20]:
mini_recs = dict((k, recs_hy[k]) for k in recs_hy.keys()[:N])
MRRs[N].append(MRR(recs=mini_recs, rel_thresh=1))
nDCGs[N].append(
nDCG(recs=mini_recs, alt_form=False, rel_thresh=False))
APs[N].append(AP_at_N(n=N, recs=recs_hy, rel_thresh=1))
Rprecs[N].append(R_precision(n_relevants=N, recs=mini_recs))
for N in [5, 10, 15, 20]:
with open('TwitterRatings/hybrid/clean/protocol.txt', 'a') as file:
file.write( "N=%s, nDCG=%s, MAP=%s, MRR=%s, R-precision=%s\n" % \
(N, mean(nDCGs[N]), mean(APs[N]), mean(MRRs[N]), mean(Rprecs[N])) )
column_names = ['userId', 'movieId', 'timestamp']
label_name = 'rating'
X, y = data[column_names].values, data[label_name].values
pos = 80000
X_train, X_test = X[:pos], X[pos:]
y_train, y_test = y[:pos], y[pos:]
# build model
from pyfm import pylibfm
model = pylibfm.FM(
num_factors=8,
num_iter=10,
validation_size=0.0,
task='regression',
reg_0=0.0,
reg_w=0.01,
reg_v=0.05,
)
self = pyfm_model_wrapper(model)
self.fit(X_train, y_train)
y_pred = self.predict(X_test)
np.mean(np.abs(y_pred - y_test))
dir(self.model)
self.model.w0
self.model.w
y_train = data.loc[data['cv'] < threshold, 'conversion']
y_test = data.loc[data['cv'] >= threshold, 'conversion']
dictTrain = list(
map(lambda ind: dict.fromkeys(getNames(train, ind), 1), train.index))
dictTest = list(
map(lambda ind: dict.fromkeys(getNames(test, ind), 1), test.index))
v = DictVectorizer()
X_train = v.fit_transform(dictTrain)
X_test = v.transform(dictTest)
# =============================================================================
# Factorization Machine
# =============================================================================
fm = pylibfm.FM(num_factors=20, num_iter=50, task="classification")
fm.fit(X_train, y_train)
# =========================================================================
# Compute MPR
# =========================================================================
data_test = pd.concat([test, y_test], axis=1)
data_test['setSize'] = list(map(len, data_test['itemSet']))
data_test = data_test.loc[(data_test['setSize'] > 1) &
(data_test['conversion'] == 1), ]
data_test['target2'] = -1
for ind in data_test.index:
data_test.loc[ind, 'target2'] = data_test.loc[ind, 'itemSet'][-1]
data_test.at[ind, 'itemSet'] = data_test.loc[ind, 'itemSet'][:-1]
'''
feature = data.iloc[:, :-1] #取特征
label = data.iloc[:, -1]
#将数组按列进行归一化
feature = minmax_scale(feature, axis=0) #此处如果不归一化，预测结果全是Nan
return feature, label
train = pd.read_csv(trainData, header=None)
test = pd.read_csv(testData, header=None)
X_train, y_train = preprocessData(train)
X_test, y_test = preprocessData(test)
X_train = [{v: k
for k, v in dict(zip(i, range(len(i)))).items()} for i in X_train]
X_test = [{v: k
for k, v in dict(zip(i, range(len(i)))).items()} for i in X_test]
v = DictVectorizer()
X_train = v.fit_transform(X_train)
X_test = v.transform(X_test)
fm = pylibfm.FM(num_factors=15,
num_iter=300,
verbose=False,
task="classification",
initial_learning_rate=0.01,
learning_rate_schedule="optimal")
fm.fit(X_train, y_train)
y_pred_label = [get_label(i) for i in fm.predict(X_test)]
print(y_pred_label)
print(accuracy_score(y_test, y_pred_label))
"age": 33
},
{
"user": "******",
"item": "20",
"age": 55
},
{
"user": "******",
"item": "10",
"age": 20
},
]
v = DictVectorizer()
X = v.fit_transform(train)
print(type(X))
print(print(v.get_feature_names()))
print(X)
# print(X.toarray())
# [[ 19. 0. 0. 0. 1. 1. 0. 0. 0.]
# [ 33. 0. 0. 1. 0. 0. 1. 0. 0.]
# [ 55. 0. 1. 0. 0. 0. 0. 1. 0.]
# [ 20. 1. 0. 0. 0. 0. 0. 0. 1.]]
y = np.repeat(1.0, X.shape[0])
print(y)
fm = pylibfm.FM()
fm.fit(X, y)
pred = fm.predict(v.transform({"user": "******", "item": "10", "age": 24}))
print(pred)
print(v.transform({"user": "******", "item": "10", "age": 24}))
print(v.transform({"user": "******", "item": "10", "age": 24}).toarray())
test = data.loc[data['cv']>=threshold,['target','itemSet']]
y_train = data.loc[data['cv']<threshold,'conversion']
y_test = data.loc[data['cv']>=threshold,'conversion']
dictTrain = list(map(lambda ind: dict.fromkeys(getNames(train,ind),1),train.index))
dictTest = list(map(lambda ind: dict.fromkeys(getNames(test,ind),1),test.index))
v = DictVectorizer()
X_train = v.fit_transform(dictTrain)
X_test = v.transform(dictTest)
# =========================================================================
# Factorization Machine
# =========================================================================
fm = pylibfm.FM(num_factors=numTraits,num_iter=100,task="classification")
fm.fit(X_train,y_train)
# =========================================================================
# Compute MPR
# =========================================================================
data_test = pd.concat([test,y_test],axis=1)
data_test = data_test.loc[data_test['conversion']==1,]
percentileRank = []
precisionAt5 = 0
precisionAt10 = 0
precisionAt20 = 0
for ind in data_test.index:
subdata = data_test.loc[ind,]
true_target = subdata['target']
def pyFM_tuning(data_path, N, with_timestamps=False, with_authors=False):
all_data, y_all, _ = loadData("eval_all_N" + str(N) + ".data",
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
v = DictVectorizer()
X_all = v.fit_transform(all_data)
defaults = {'f': 100, 'mi': 20, 'bias': True, 'oneway': True , 'init_stdev': 0.1, 'val_size': 0.01, 'lr_s': 'optimal', 'lr': 0.01, \
'invscale_pow': 0.5, 'optimal_denom': 0.001, 'shuffle': True, 'seed': 28} #cambio del original: f:20, mi:1
results = dict((param, {}) for param in defaults.keys())
for param in [
'mi', 'f', 'bias', 'oneway', 'init_stdev', 'val_size', 'lr_s',
'lr', 'invscale_pow', 'optimal_denom', 'shuffle', 'seed'
]:
if param == 'mi':
for i in [1, 5, 10, 20, 50, 100, 150, 200]:
defaults['mi'] = i
results['mi'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['mi'] = opt_value(results=results['mi'], metric='rmse')
elif param == 'f':
for i in range(20, 2020, 20):
defaults['f'] = i
results['f'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['f'] = opt_value(results=results['f'], metric='rmse')
elif param == 'bias':
for i in [True, False]:
defaults['bias'] = i
results['bias'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['bias'] = opt_value(results=results['bias'],
metric='rmse')
elif param == 'oneway':
for i in [True, False]:
defaults['oneway'] = i
results['oneway'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['oneway'] = opt_value(results=results['oneway'],
metric='rmse')
elif param == 'init_stdev':
for i in [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5, 1.0]:
defaults['init_stdev'] = i
results['init_stdev'][i] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['init_stdev'] = opt_value(results=results['init_stdev'],
metric='rmse')
elif param == 'val_size':
for i in [0.001, 0.01, 0.1, 0.5, 0.8, 0.9]:
defaults['val_size'] = i
results['val_size'][i] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['val_size'] = opt_value(results=results['val_size'],
metric='rmse')
elif param == 'lr_s':
for i in ['constant', 'optimal', 'invscaling']:
defaults['lr_s'] = i
if i == 'optimal':
for j in [
0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05, 0.1, 0.5
]:
defaults['optimal_denom'] = j
results['optimal_denom'][j] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['optimal_denom'] = opt_value(
results=results['optimal_denom'], metric='rmse')
results['lr_s'][i] = results['optimal_denom'][
defaults['optimal_denom']]
elif i == 'invscaling':
for j in [0.001, 0.05, 0.1, 0.5, 0.8, 1.0]:
defaults['invscale_pow'] = j
results['invscale_pow'][j] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['invscale_pow'] = opt_value(
results=results['invscale_pow'], metric='rmse')
results['lr_s'][i] = results['invscale_pow'][
defaults['invscale_pow']]
elif i == 'constant':
results['lr_s'][i] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['lr_s'] = opt_value(results=results['lr_s'],
metric='rmse')
elif param == 'lr':
for i in [0.001, 0.003, 0.005, 0.01, 0.02, 0.03, 0.04,
0.05]: #0.07, 0.08, 0.1]:
defaults['lr'] = i
results['lr'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['lr'] = opt_value(results=results['lr'], metric='rmse')
elif param == 'shuffle':
for i in [True, False]:
defaults['shuffle'] = i
results['shuffle'][i] = pyFMJob(
data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['shuffle'] = opt_value(results=results['shuffle'],
metric='rmse')
elif param == 'seed':
for i in [10, 20, 28, 30, 50]:
defaults['seed'] = i
results['seed'][i] = pyFMJob(data_path=data_path,
params=defaults,
N=N,
vectorizer=v,
with_timestamps=with_timestamps,
with_authors=with_authors)
defaults['seed'] = opt_value(results=results['seed'],
metric='rmse')
# Real testing
train_data, y_tr, _ = loadData('eval_train_N' + str(N) + '.data',
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
X_tr = v.transform(train_data)
fm = pylibfm.FM(num_factors=defaults['f'], num_iter=defaults['mi'], k0=defaults['bias'], k1=defaults['oneway'], init_stdev=defaults['init_stdev'], \
validation_size=defaults['val_size'], learning_rate_schedule=defaults['lr_s'], initial_learning_rate=defaults['lr'], \
power_t=defaults['invscale_pow'], t0=defaults['optimal_denom'], shuffle_training=defaults['shuffle'], seed=defaults['seed'], \
task='regression', verbose=True)
fm.fit(X_tr, y_tr)
test_data, y_te, _ = loadData('test/test_N' + str(N) + '.data',
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
X_te = v.transform(test_data)
preds = fm.predict(X_te)
rmse = sqrt(mean_squared_error(y_te, preds))
print("FM RMSE: %.4f" % rmse)
with open(
'TwitterRatings/pyFM/opt_params_tmstmp' + str(with_timestamps) +
'_auth' + str(with_authors) + '.txt', 'w') as f:
for param in defaults:
f.write("{param}:{value}\n".format(param=param,
value=defaults[param]))
f.write("RMSE:{rmse}".format(rmse=rmse))
with open(
'TwitterRatings/pyFM/params_rmses_tmstmp' + str(with_timestamps) +
'_auth' + str(with_authors) + '.txt', 'w') as f:
for param in results:
for value in results[param]:
f.write("{param}={value}\t : {RMSE}\n".format(
param=param, value=value, RMSE=results[param][value]))
return defaults
def pyFM_protocol_evaluation(data_path,
params,
with_timestamps=False,
with_authors=False):
solr = "http://localhost:8983/solr/grrecsys"
# userId = '33120270'
all_data, y_all, items = loadData("eval_all_N20.data",
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
v = DictVectorizer()
X_all = v.fit_transform(all_data)
test_c = consumption(
ratings_path='TwitterRatings/funkSVD/data/test/test_N20.data',
rel_thresh=0,
with_ratings=True)
train_c = consumption(
ratings_path='TwitterRatings/funkSVD/data/eval_train_N20.data',
rel_thresh=0,
with_ratings=False)
MRRs = dict((N, []) for N in [5, 10, 15, 20])
nDCGs = dict((N, []) for N in [5, 10, 15, 20])
APs = dict((N, []) for N in [5, 10, 15, 20])
Rprecs = dict((N, []) for N in [5, 10, 15, 20])
train_data, y_tr, _ = loadData('eval_train_N20.data',
data_path=data_path,
with_timestamps=with_timestamps,
with_authors=with_authors)
X_tr = v.transform(train_data)
fm = pylibfm.FM(num_factors=params['f'], num_iter=params['mi'], k0=params['bias'], k1=params['oneway'], init_stdev=params['init_stdev'], \
validation_size=params['val_size'], learning_rate_schedule=params['lr_s'], initial_learning_rate=params['lr'], \
power_t=params['invscale_pow'], t0=params['optimal_denom'], shuffle_training=params['shuffle'], seed=params['seed'], \
task='regression', verbose=True)
fm.fit(X_tr, y_tr)
p = 0
for userId in test_c:
logging.info("#u: {0}/{1}".format(p, len(test_c)))
p += 1
user_rows = [{
'user_id': str(userId),
'item_id': str(itemId)
} for itemId in items]
X_te = v.transform(user_rows)
preds = fm.predict(X_te)
book_recs = [
itemId for _, itemId in sorted(zip(preds, items), reverse=True)
]
book_recs = remove_consumed(user_consumption=train_c[userId],
rec_list=book_recs)
book_recs = recs_cleaner(solr=solr,
consumpt=train_c[userId],
recs=book_recs[:100])
recs = user_ranked_recs(user_recs=book_recs,
user_consumpt=test_c[userId])
for N in [5, 10, 15, 20]:
mini_recs = dict((k, recs[k]) for k in recs.keys()[:N])
MRRs[N].append(MRR(recs=mini_recs, rel_thresh=1))
nDCGs[N].append(
nDCG(recs=mini_recs, alt_form=False, rel_thresh=False))
APs[N].append(AP_at_N(n=N, recs=recs, rel_thresh=1))
Rprecs[N].append(R_precision(n_relevants=N, recs=mini_recs))
for N in [5, 10, 15, 20]:
with open(
'TwitterRatings/pyFM/clean/protocol_tmstmp' +
str(with_timestamps) + '_auth' + str(with_authors) + '.txt',
'a') as file:
file.write( "N=%s, nDCG=%s, MAP=%s, MRR=%s, R-precision=%s\n" % \
(N, mean(nDCGs[N]), mean(APs[N]), mean(MRRs[N]), mean(Rprecs[N])) )
