def test_fm_sgr_ranking():
w0, w, V, y, X = get_test_problem()
X_test = X.copy()
X_train = X.copy()
import itertools
pairs = [p for p in itertools.combinations(range(len(y)), 2)]
compares = np.zeros((len(pairs), 2), dtype=np.float64)
for i, p in enumerate(pairs):
if y[p[0]] > y[p[1]]:
compares[i, 0] = p[0]
compares[i, 1] = p[1]
else:
compares[i, 0] = p[1]
compares[i, 1] = p[0]
print(compares)
fm = bpr.FMRecommender(n_iter=2000,
init_stdev=0.01, l2_reg_w=.5, l2_reg_V=.5, rank=2,
step_size=.002, random_state=11)
fm.fit(X_train, compares)
y_pred = fm.predict(X_test)
y_pred = np.argsort(y_pred)
print(y)
print(y_pred)
print(np.argsort(y))
assert utils.kendall_tau(np.argsort(y), y_pred) == 1
def __init__(self, n_iter=100,
init_stdev=0.1,
rank=8,
random_state=123,
l2_reg_w=0.1,
l2_reg_V=0.1,
l2_reg=0,
step_size=0.1):
self.fastFM = bpr.FMRecommender(n_iter=n_iter,
init_stdev=init_stdev,
rank=rank,
random_state=random_state,
l2_reg_w=l2_reg_w,
l2_reg_V=l2_reg_V,
l2_reg=l2_reg,
step_size=step_size)
self.URM = None
self.ICM = None
self.train_x = None
self.train_y = None
self.test_playlists = None
self.test_tracks = None
self.test_x = None
self.train_length = 222689
self.non_zeros = 0
self.n_playlists = 45649
self.n_tracks = 100000
self.n_attributes = 77040
def train_model(x_train,
y_train,
n_iter,
init_stdev=0.1,
rank=2,
l2_reg_w=0.1,
l2_reg_V=0.5):
fm = bpr.FMRecommender(n_iter=n_iter,
init_stdev=init_stdev,
rank=rank,
l2_reg_w=l2_reg_w,
l2_reg_V=l2_reg_V)
fm.fit(x_train, y_train)
return fm
def fastFMJob_bpr(data_path, params, N, vectorizer, train_sets, pairs, items):
ndcgs = []
logging.info("Evaluando con params: {0}".format(params))
for i in range(1, 4 + 1):
fm = bpr.FMRecommender(n_iter=params['mi'], init_stdev=params['init_stdev'], rank=params['f'], random_state=123, \
l2_reg_w=params['l2_reg_w'], l2_reg_V=params['l2_reg_V'], l2_reg=params['l2_reg'], step_size=params['step_size'])
X_tr = vectorizer.transform(train_sets[i])
fm.fit(X_tr, pairs[i])
val_c = consumption(
ratings_path='TwitterRatings/funkSVD/data/val/val_N' + str(N) +
'.' + str(i),
rel_thresh=0,
with_ratings=True)
train_c = consumption(
ratings_path='TwitterRatings/funkSVD/data/train/train_N' + str(N) +
'.' + str(i),
rel_thresh=0,
with_ratings=True)
users_ndcgs = []
for userId in val_c:
val_data, y_va, _ = loadData_bpr('val/val_N' + str(N) + '.' +
str(i),
data_path=data_path,
test=True,
userId_va=userId)
X_va = vectorizer.transform(val_data)
preds = fm.predict(X_va)
preds = np.argsort(-preds)
users_ndcgs.append(
ndcg_bpr(preds=preds,
vectorizer=vectorizer,
matrix=X_va,
user_data=train_c[userId],
user_val=val_c[userId],
N=N))
fold_ndcg = mean(users_ndcgs)
logging.info("FM fold {0} nDCG: {1}. Solver: BPR".format(i, fold_ndcg))
ndcgs.append(fold_ndcg)
return mean(ndcgs)
def fastFM_protocol_evaluation_bpr(data_path, params):
# userId = '33120270'
solr = "http://localhost:8983/solr/grrecsys"
all_data, y_all, items = loadData_bpr("eval_all_N20.data",
data_path=data_path)
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)
all_c = consumption(
ratings_path='TwitterRatings/funkSVD/data/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])
train_data, y_tr, _ = loadData_bpr('eval_train_N20.data',
data_path=data_path)
X_tr = v.transform(train_data)
fm = bpr.FMRecommender(n_iter=params['mi'], init_stdev=params['init_stdev'], rank=params['f'], random_state=123, \
l2_reg_w=params['l2_reg_w'], l2_reg_V=params['l2_reg_V'], l2_reg=params['l2_reg'], step_size=params['step_size'])
pairs_tr = make_pairs(X_tr, y_tr)
fm.fit(X_tr, pairs_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)
preds = np.argsort(-preds) # ordenamos las predicciones
book_recs = []
for i in range(len(preds)):
print("i={}".format(i))
pred_row = preds[i]
l = v.inverse_transform(X_te[pred_row, :])[0].keys()
pred_itemId = [s for s in l if "item" in s][0].split('=')[-1]
book_recs.append(pred_itemId)
if i == 100: break # no necesitamos más de 100
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/fastFM/bpr/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])) )
np.random.randint(0, nItem, sap)] = 1
beginY = sap * i
Y_train[beginY:beginY + sap, 0] = begin
Y_train[beginY:beginY + sap, 1] = np.arange(begin + 1, begin + sap + 1)
if i % 1000 == 0:
print("Constructing %d/%d" % (i, nTrain))
X_train = hstack([X_ItemTrain, X_TargetTrain]).tocsc()
# Build Model
print('Start training')
embed_len = 20
fm = bpr.FMRecommender(n_iter=5000000,
init_stdev=0.1,
rank=embed_len,
l2_reg_w=0,
l2_reg_V=0,
l2_reg=0,
step_size=0.1)
fm.fit(X_train, Y_train)
print('Start evaluation')
X_Target = eye(nItem + 1, dtype=np.float32)
atK = np.arange(5, 51, 5)
ranklist = []
nTest = len(testLast)
recs = []
for i, row in enumerate(testLast):
# X_User = lil_matrix((nItem + 1, nUser + 1), dtype=np.float32)
