def createRecSystem(self):
db, idb, movies = dataload.read_data_to_hash(self.trainset)
self.rec = KNeighborRegressor(db,
idb,
k=self.k,
sim_method=self.sim_method,
rec_type=self.rec_type)
def createRecSystem(self):
db, idb, movies = dataload.read_data_to_hash(self.trainset)
self.rec = KNeighborRegressor(db, idb, k=self.k,
sim_method=self.sim_method, rec_type=self.rec_type)
class Evaluater(object):
def __init__(self, datafile, rec_type='ub', eval_metric='mae',
k=5, sim_method=sim_cosine, test_percentage=30):
self.datafile = datafile
self.per = test_percentage / 100
self.k = k
self.sim_method = sim_method
self.eval_metric = eval_metric
self.rec_type = rec_type
# These will be initiated later.
self.dataset = None
self.n_inst = None
self.trainset = None
self.testset = None
self.rec = None
#FIXME
self.prepareEvaluater()
print self
print '\n'
def prepareEvaluater(self):
self.dataset = dataload.get_dataset(self.datafile)
#if self.rec_type == 'ub': # userbased
#self.dataset = dataload.get_dataset(self.datafile)
#elif self.rec_type == 'ib': # itembased
#pass
#else:
#stderr.write("Please enter ub [user-based] or ib [item-based]")
#exit(1)
#if self.dataset:
self.__prepare_datasets()
self.createRecSystem()
def createRecSystem(self):
db, idb, movies = dataload.read_data_to_hash(self.trainset)
self.rec = KNeighborRegressor(db, idb, k=self.k,
sim_method=self.sim_method, rec_type=self.rec_type)
def __prepare_datasets(self):
#TODO this method may be in dataload module.
self.n_inst = len(self.dataset)
shuffle(self.dataset)
n_test_inst = int(self.per * self.n_inst)
self.testset = self.dataset[:n_test_inst]
self.trainset = self.dataset[n_test_inst:]
total = sum(map(len, [self.trainset, self.testset]))
assert total == self.n_inst
def __calc_mae(self, test_db):
""" This method evaluates that how much our prediction's
are accurate.
This distinction is very important. MAE is able to characterize the
accuracy of prediction not accuracy of recommendation[1].
If you want to evaluate accuracy of prediction, you should
use 'pr' rather than 'mae'. 'pr' provides you Precision and Recall.
[1]: Symeonidis, P. et al., 2006. Collaborative Filtering: Fallacies
and Insights in Measuring Similarity. In citeulikeorg. Citeseer,
p. 56. Available at:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.61.8340&
rep=rep1&type=pdf.
"""
rating_list = []
for user, itemRating in test_db.iteritems():
# user : integer - user_id
# itemRating : dictionary - { item_id:Rating, ... }
rating = self.rec.predict(user, itemRating.keys()[0])
rating_list.append([rating, itemRating.values()[0]])
rating_array = np.array(rating_list)
print "Number of key error: %s" % self.rec.adapt.numberOfKeyError
#TODO: self.eval_metric?
return mae(rating_array[:,0], rating_array[:,1])
def __calc_pr(self, test_db, k=1, N=20, I=300):
#TODO args, kwargs
#pr = PrecisionRecall(self.dataset, test_db, N, I)
pr = PrecisionRecall(self.trainset, test_db, N, I)
rating_list = []
actual_testset = pr.get_important_items_from_testset() # This item is high rated by users.
#print 'Actual Length of the Testset%s\n' % actual_testset
for user, item in actual_testset:
#print "\nNew test for user_id=%s, item_id=%s\n" % (user, item)
unrev_items = pr.get_unrelevant_items(user)
if unrev_items is not None:
unrev_items.append(item) # added our relevant item for prediction
assert len(unrev_items) == I + 1
for item in unrev_items:
key = str(user) + '_' + str(item)
rating = self.rec.predict(user, item, signature=key)
rating_list.append([item, rating])
topNList = pr.create_topN_list(rating_list)
pr.use_for_pr(topNList, item) # use this inf. for calculating PR
precision, recall = pr.evaluate_pr()
return precision, recall
def evaluate(self, k=1, N=20, I=300):
"""
Input:
-> k: Number of item(s) which are/is the user appreciates
most. (Only for Precision & Recall)
"""
test_db, test_idb, m = dataload.read_data_to_hash(self.testset)
if self.rec_type == 'ib':
test_db, test_idb = test_idb, test_db
if self.eval_metric == 'mae':
return self.__calc_mae(test_db)
elif self.eval_metric == 'pr': # precision & Recall
return self.__calc_pr((test_db, test_idb), k, N, I)
else:
stderr.write("There is no such evaluation metric you can use")
exit(1)
def __str__(self):
p = self.per
d_f = self.datafile
k = self.k
sim = self.sim_method.func_name
return 'Evaluater\nDatafile:%s\ntest_percentage:%s\nk:%s\nSimilarity Method:%s' % (d_f,p,k,sim)
class Evaluater(object):
def __init__(self,
datafile,
rec_type='ub',
eval_metric='mae',
k=5,
sim_method=sim_cosine,
test_percentage=30):
self.datafile = datafile
self.per = test_percentage / 100
self.k = k
self.sim_method = sim_method
self.eval_metric = eval_metric
self.rec_type = rec_type
# These will be initiated later.
self.dataset = None
self.n_inst = None
self.trainset = None
self.testset = None
self.rec = None
#FIXME
self.prepareEvaluater()
print self
print '\n'
def prepareEvaluater(self):
self.dataset = dataload.get_dataset(self.datafile)
#if self.rec_type == 'ub': # userbased
#self.dataset = dataload.get_dataset(self.datafile)
#elif self.rec_type == 'ib': # itembased
#pass
#else:
#stderr.write("Please enter ub [user-based] or ib [item-based]")
#exit(1)
#if self.dataset:
self.__prepare_datasets()
self.createRecSystem()
def createRecSystem(self):
db, idb, movies = dataload.read_data_to_hash(self.trainset)
self.rec = KNeighborRegressor(db,
idb,
k=self.k,
sim_method=self.sim_method,
rec_type=self.rec_type)
def __prepare_datasets(self):
#TODO this method may be in dataload module.
self.n_inst = len(self.dataset)
shuffle(self.dataset)
n_test_inst = int(self.per * self.n_inst)
self.testset = self.dataset[:n_test_inst]
self.trainset = self.dataset[n_test_inst:]
total = sum(map(len, [self.trainset, self.testset]))
assert total == self.n_inst
def __calc_mae(self, test_db):
""" This method evaluates that how much our prediction's
are accurate.
This distinction is very important. MAE is able to characterize the
accuracy of prediction not accuracy of recommendation[1].
If you want to evaluate accuracy of prediction, you should
use 'pr' rather than 'mae'. 'pr' provides you Precision and Recall.
[1]: Symeonidis, P. et al., 2006. Collaborative Filtering: Fallacies
and Insights in Measuring Similarity. In citeulikeorg. Citeseer,
p. 56. Available at:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.61.8340&
rep=rep1&type=pdf.
"""
rating_list = []
for user, itemRating in test_db.iteritems():
# user : integer - user_id
# itemRating : dictionary - { item_id:Rating, ... }
rating = self.rec.predict(user, itemRating.keys()[0])
rating_list.append([rating, itemRating.values()[0]])
rating_array = np.array(rating_list)
print "Number of key error: %s" % self.rec.adapt.numberOfKeyError
#TODO: self.eval_metric?
return mae(rating_array[:, 0], rating_array[:, 1])
def __calc_pr(self, test_db, k=1, N=20, I=300):
#TODO args, kwargs
#pr = PrecisionRecall(self.dataset, test_db, N, I)
pr = PrecisionRecall(self.trainset, test_db, N, I)
rating_list = []
actual_testset = pr.get_important_items_from_testset(
) # This item is high rated by users.
#print 'Actual Length of the Testset%s\n' % actual_testset
for user, item in actual_testset:
#print "\nNew test for user_id=%s, item_id=%s\n" % (user, item)
unrev_items = pr.get_unrelevant_items(user)
if unrev_items is not None:
unrev_items.append(
item) # added our relevant item for prediction
assert len(unrev_items) == I + 1
for item in unrev_items:
key = str(user) + '_' + str(item)
rating = self.rec.predict(user, item, signature=key)
rating_list.append([item, rating])
topNList = pr.create_topN_list(rating_list)
pr.use_for_pr(topNList,
item) # use this inf. for calculating PR
precision, recall = pr.evaluate_pr()
return precision, recall
def evaluate(self, k=1, N=20, I=300):
"""
Input:
-> k: Number of item(s) which are/is the user appreciates
most. (Only for Precision & Recall)
"""
test_db, test_idb, m = dataload.read_data_to_hash(self.testset)
if self.rec_type == 'ib':
test_db, test_idb = test_idb, test_db
if self.eval_metric == 'mae':
return self.__calc_mae(test_db)
elif self.eval_metric == 'pr': # precision & Recall
return self.__calc_pr((test_db, test_idb), k, N, I)
else:
stderr.write("There is no such evaluation metric you can use")
exit(1)
def __str__(self):
p = self.per
d_f = self.datafile
k = self.k
sim = self.sim_method.func_name
return 'Evaluater\nDatafile:%s\ntest_percentage:%s\nk:%s\nSimilarity Method:%s' % (
d_f, p, k, sim)