def algo_tester(data_object):
'''
Produces a dataframe displaying all the different RMSE's, test & train times of the different surprise algorithms
---Parameters---
data_object(variable) created from the read_data_surprise function
---Returns---
returns a dataframe where you can compare the performance of different algorithms
'''
benchmark = []
algos = [
sp.SVDpp(),
sp.SVD(),
sp.SlopeOne(),
sp.NMF(),
sp.NormalPredictor(),
sp.KNNBaseline(),
sp.KNNBasic(),
sp.KNNWithMeans(),
sp.KNNWithZScore(),
sp.BaselineOnly(),
sp.CoClustering()
]
# Iterate over all algorithms
for algorithm in algos:
# Perform cross validation
results = cross_validate(algorithm,
data_object,
measures=['RMSE'],
cv=3,
verbose=False)
# Get results & append algorithm name
tmp = pd.DataFrame.from_dict(results).mean(axis=0)
tmp = tmp.append(
pd.Series([str(algorithm).split(' ')[0].split('.')[-1]],
index=['Algorithm']))
benchmark.append(tmp)
benchmark = pd.DataFrame(benchmark).set_index('Algorithm').sort_values(
'test_rmse')
return benchmark
def model_fit(self):
'''
Train model using surprise.SVD algorithm.
'''
self.build_trainset()
algo = self._algo_choise
if algo == 'SVD':
self.algorithm = surprise.SVD()
elif algo == 'Baseline':
self.algorithm = surprise.BaselineOnly()
elif algo == 'SlopeOne':
self.algorithm = surprise.SlopeOne()
elif algo == 'CoClustering':
self.algorithm = surprise.CoClustering()
else:
self.algorithm = surprise.KNNBasic()
print('Training Recommender System using %s...' % algo)
self.algorithm.fit(self.trainset)
self.ratings_changed = False
print('Done')
epochs=2,
validation_split=0.1,
shuffle=True)
y_pred = model.predict([df_hybrid_test['User'], df_hybrid_test['Movie'], test_tfidf])
y_true = df_hybrid_test['Rating'].values
rmse = np.sqrt(mean_squared_error(y_pred=y_pred, y_true=y_true))
print('\n\nTesting Result With Keras Hybrid Deep Learning: {:.4f} RMSE'.format(rmse))
# Load dataset into surprise specific data-structure
data = sp.Dataset.load_from_df(df_filterd[['User', 'Movie', 'Rating']].sample(20000), sp.Reader())
benchmark = []
# Iterate over all algorithms
for algorithm in [sp.SVD(), sp.SVDpp(), sp.SlopeOne(), sp.NMF(), sp.NormalPredictor(), sp.KNNBaseline(), sp.KNNBasic(), sp.KNNWithMeans(), sp.KNNWithZScore(), sp.BaselineOnly(), sp.CoClustering()]:
# Perform cross validation
results = cross_validate(algorithm, data, measures=['RMSE', 'MAE'], cv=3, verbose=False)
# Get results & append algorithm name
tmp = pd.DataFrame.from_dict(results).mean(axis=0)
tmp = tmp.append(pd.Series([str(algorithm).split(' ')[0].split('.')[-1]], index=['Algorithm']))
# Store data
benchmark.append(tmp)
# Store results
surprise_results = pd.DataFrame(benchmark).set_index('Algorithm').sort_values('test_rmse', ascending=False)
# Get data
data = surprise_results[['test_rmse', 'test_mae']]
def main(train_df, target_df, cache_name="test", force_recompute=[]):
"""Train multiple models on train_df and predicts target_df
Predictions are cached. If the indices don't match the indices of
target_df, the cache is discarded.
By default, if a method was already computed it is not recomputed again
(except if the method name is listed in force_recompute). cache_name
is the name to use to read and write the cache.
Arguments:
train_df {dataframe} -- Training dataframe
target_df {dataframe} -- Testing dataframe
Keyword Arguments:
cache_name {str} -- Name to use for caching (default: {"test"})
force_recompute {list} -- Name(s) of methods to recompute, whether or
not it was already computed. Useful to only recompute single methods
without discarding the rest. (default: {[]})
Returns:
Dataframe -- Dataframe with predictions for each methods as columns,
IDs as indices
"""
global algo_in_use
CACHED_DF_FILENAME = os.path.dirname(
os.path.abspath(__file__)) +\
"/cache/cached_predictions_{}.pkl".format(cache_name)
train_df = preprocess_df(train_df)
trainset = pandas_to_data(train_df)
ids_to_predict = target_df["Id"].to_list()
# try to retrieve backup dataframe
try:
print("Retrieving cached predictions")
all_algos_preds_df = pd.read_pickle(CACHED_DF_FILENAME)
print("Ensuring cached IDs match given IDs")
assert sorted(ids_to_predict) == sorted(
all_algos_preds_df.index.values)
print("Indices match, continuing")
except (FileNotFoundError, AssertionError):
print("No valid cached predictions found")
all_algos_preds_df = pd.DataFrame(ids_to_predict, columns=["Id"])
all_algos_preds_df.set_index("Id", inplace=True)
all_algos = {
"SVD": spr.SVD(n_factors=200, n_epochs=100),
"Baseline": spr.BaselineOnly(),
"NMF": spr.NMF(n_factors=30, n_epochs=100),
"Slope One": spr.SlopeOne(),
"KNN Basic": spr.KNNBasic(k=60),
"KNN Means": spr.KNNWithMeans(k=60),
"KNN Baseline": spr.KNNBaseline(),
"KNN Zscore": spr.KNNWithZScore(k=60),
"SVD ++": spr.SVDpp(n_factors=40, n_epochs=100),
"Co Clustering": spr.CoClustering()
}
for name in all_algos:
print("##### {} ####".format(name))
if name in force_recompute and name in all_algos_preds_df.columns:
all_algos_preds_df.drop(name, axis=1, inplace=True)
if name in all_algos_preds_df.columns:
print("Already computed {}, skipping".format(name))
continue
algo = all_algos[name]
time.sleep(1)
algo.fit(trainset)
time.sleep(1)
algo_in_use = algo
print("Generating predictions...")
predictions = parallelize_predictions(ids_to_predict, 80)
print("Done. Merging with previous results")
this_algo_preds_df = pd.DataFrame(predictions, columns=["Id", name])
this_algo_preds_df.set_index("Id", inplace=True)
all_algos_preds_df = pd.merge(all_algos_preds_df,
this_algo_preds_df,
left_index=True,
right_index=True)
all_algos_preds_df.to_pickle(CACHED_DF_FILENAME)
print("DONE computing surprize")
return all_algos_preds_df
def build_network(self):
self.model = surprise.CoClustering()
dfs_path = os.path.join(root_dir, 'Data/datasets/')
ratings_df = pd.read_csv(os.path.join(dfs_path, 'ratings_expl.csv'),
sep=';',
encoding='latin-1',
low_memory=False)
# 加载数据
reader = surprise.Reader(rating_scale=(1, 10))
data = surprise.Dataset.load_from_df(ratings_df[['user_id', 'isbn', 'rating']],
reader)
kf = KFold(n_splits=5)
# data = Dataset.load_builtin('jester')
# Kfold
algo1 = SVD()
algo2 = surprise.BaselineOnly()
algo3 = surprise.KNNBasic()
algo4 = surprise.CoClustering()
for trainset, testset in kf.split(data):
# SVD
algo1.fit(trainset)
pSVD = algo1.test(testset)
# 计算并打印RMSE
print("SVD-")
accuracy.rmse(pSVD, verbose=True)
#Baseline
algo2.fit(trainset)
pBase = algo2.test(testset)
print("BaseLine-")
accuracy.rmse(pBase)
#Baseline
algo3.fit(trainset)
pknn = algo3.test(testset)
# identifier_df_test = X_test[['user_id', 'business_id']]
# A reader is needed but only the rating_scale param is requiered.
reader = Reader(rating_scale=(1, 5))
data = Dataset.load_from_df(train_df, reader)
sim_options = {'user_based': [False]}
results = []
# Iterate over all algorithms
for algorithm in [
SVD(),
surprise.NMF(),
surprise.SlopeOne(),
surprise.CoClustering(),
surprise.KNNBasic(sim_options=sim_options),
surprise.KNNWithMeans(sim_options=sim_options),
surprise.KNNWithZScore(sim_options=sim_options),
surprise.KNNBaseline(sim_options=sim_options),
surprise.NormalPredictor(),
surprise.BaselineOnly()
]:
# Get string of algname for naming a pickle file a useful name
alg_name = str(algorithm)
alg_name = alg_name[alg_name.find('.') + 1:]
alg_name = alg_name[alg_name.find('.') + 1:]
alg_name = alg_name[alg_name.find('.') + 1:]
alg_name = alg_name[:alg_name.find('object') - 1]
# defining the number of folds = 5
print("Performing splits...")
kf = sp.model_selection.KFold(n_splits=5, random_state=0)
print("Done.")
###
### PART 1.1
###
'''
application of all algorithms for recommendation made available by
“Surprise” libraries, according to their default configuration.
'''
algorithms = [sp.NormalPredictor(), sp.BaselineOnly(), sp.KNNBasic(),\
sp.KNNWithMeans(), sp.KNNWithZScore(), sp.KNNBaseline(),\
sp.SVD(), sp.SVDpp(), sp.NMF(), sp.SlopeOne(), sp.CoClustering()]
for elem in algorithms:
start_time = time.time()
algo = elem
sp.model_selection.cross_validate(algo, data, measures=['RMSE'], \
cv=kf, n_jobs = 2, verbose=True)
print("--- %s seconds ---" % (time.time() - start_time))
print()
###
### PART 1.2
###
'''
Improvement of the quality of both KNNBaseline and SVD methods,
by performing hyper-parameters tuning over 5-folds
Random-Search-Cross-Validation - KNN
def SurpriseBased(table, relation_name, parameters, verbose=False):
"""
"""
report = {}
# Initial checks
param_keys = [k for k, v in parameters.items()]
if ('max_scale' not in param_keys) or ('min_scale' not in param_keys):
raise ValueError(
'max_scale and min_scale must be specified in parameters for explicit RS.'
)
if 'model_size' not in param_keys:
raise ValueError(
'model_size must be specified in parameters for SURPRISE-based RS.'
)
if 'topK_predictions' not in param_keys:
raise ValueError(
'A size (K) must be given for the recommended list size (topK).')
# Retrieving names
start_group = table.start_group.iloc[0]
end_group = table.end_group.iloc[0]
timestamp = pd.Timestamp('')
# Retrieving the table of the bipartite graph in SURPRISE format
table = table[['start_object', 'end_object', 'value']]
reader = surprise.Reader(rating_scale=(parameters['min_scale'],
parameters['max_scale']))
data = surprise.Dataset.load_from_df(table, reader)
# Selecting the method from the SURPRISE module
if parameters['method'] == 'UBCF':
method = surprise.KNNBasic(k=parameters['model_size'], verbose=verbose)
elif parameters['method'] == 'Z-UBCF':
method = surprise.KNNWithZScore(k=parameters['model_size'])
elif parameters['method'] == 'IBCF':
method = surprise.KNNBasic(k=parameters['model_size'],
sim_options={'user_based': False})
elif parameters['method'] == 'SVD':
method = surprise.SVD(n_factors=parameters['model_size'])
elif parameters['method'] == 'NMF':
method = surprise.NMF(n_factors=parameters['model_size'])
elif parameters['method'] == 'CClustering':
method = surprise.CoClustering(n_cltr_u=parameters['model_size'],
n_cltr_i=parameters['model_size'])
else:
raise ValueError('Unrecognized SURPRISE-based RS method named %s' %
parameters['method'])
# Computing utility metrics if so specified
if 'RMSE' in param_keys:
if parameters['RMSE']:
results = surprise.model_selection.validation.cross_validate(
method, data, measures=['rmse'], cv=5, verbose=verbose)
rmse = results['test_rmse'].mean()
report['RMSE'] = rmse
# Training the prediction method
trainset = data.build_full_trainset()
del data
method.fit(trainset)
# Retrieving unobserved pairs
t = TCounter()
VerboseMessage(verbose, 'Producing unobserved links...')
unobserved_links = trainset.build_anti_testset()
VerboseMessage(
verbose,
'Unobserved links produced in %s.' % (ETSec2ETTime(TCounter() - t)))
# Making the predictions
t = TCounter()
VerboseMessage(verbose, 'Making predictions for unobserved links...')
predictions = method.test(unobserved_links)
VerboseMessage(
verbose, 'Predictions for Unobserved links produced in %s.' %
(ETSec2ETTime(TCounter() - t)))
# Prefiltering predictions with lower scores
if 'prefilter_score' in param_keys:
t = TCounter()
VerboseMessage(
verbose, 'Prefiltering %d predictions scores lower than %0.1f...' %
(len(predictions), parameters['prefilter_threshold']))
predictions = [
p for p in predictions if p[3] > parameters['prefilter_threshold']
]
VerboseMessage(
verbose, 'Predictions prefiltered in %s, %d remaining.' %
(ETSec2ETTime(TCounter() - t), len(predictions)))
# Selecting only top K predictions
t = TCounter()
VerboseMessage(
verbose,
'Selecting top %d predictions...' % (parameters['topK_predictions']))
top_recs = defaultdict(list)
for uid, iid, true_r, est, _ in predictions:
top_recs[uid].append((iid, est))
for uid, user_ratings in top_recs.items():
user_ratings.sort(key=lambda x: x[1], reverse=True)
top_recs[uid] = user_ratings[:parameters['topK_predictions']]
VerboseMessage(
verbose,
'Predictions selected in %s.' % (ETSec2ETTime(TCounter() - t)))
# Putting the predictions in a DataFrame
predictions_table = pd.DataFrame(columns=[
'relation', 'start_group', 'start_object', 'end_group', 'end_object',
'value', 'timestamp'
])
counter = 0
t = TCounter()
VerboseMessage(verbose, 'Arranging predictions into a DataFrame table...')
for k, v in top_recs.items():
for r in v:
predictions_table.loc[counter] = [
relation_name, start_group, k, end_group, r[0], r[1], timestamp
]
counter += 1
VerboseMessage(
verbose, 'Predictions arranged into a table in %s.' %
(ETSec2ETTime(TCounter() - t)))
return predictions_table, report
elif args.model == 'KNNBasic':
model = surprise.KNNBasic()
elif args.model == 'KNNWithMeans':
model = surprise.KNNWithMeans()
elif args.model == 'KNNWithZScore':
model = surprise.KNNWithZScore()
elif args.model == 'KNNBaseline':
model = surprise.KNNBaseline()
elif args.model == 'SVD':
model = surprise.SVD()
elif args.model == 'SVDpp':
model = surprise.SVDpp(verbose=True)
elif args.model == 'NMF':
model = surprise.NMF()
elif args.model == 'SlopeOne':
model = surprise.SlopeOne()
elif args.model == 'CoClustering':
model = surprise.CoClustering()
# cross_validate(model, trainset, cv=5, verbose=True)
model.fit(trainset)
lines = []
test_path = path + '/Data/test_format.txt'
for line in tqdm(open(test_path, 'r').readlines()):
user_id, item_id, timestamp, *tags = line.strip().split(',')
rating = model.predict(user_id, item_id).est
lines.append("{:.5}\n".format(float(rating)))
open(path + '/Data/submit.txt', 'w').writelines(lines)