def test_dynamic_updates(self):
"""
TensorCoFi dynamic update
We will take a tensor cofi. Train the model, evaluate it. Then we remove all the user factors
and recompute them using the online_user_factors to check if the performance is almost the same...
"""
pyTF = PyTensorCoFi(n_factors=20, n_iterations=5, c_lambda=0.05, c_alpha=40)
evaluator = Evaluator()
tf = TensorCoFi(n_factors=2, n_iterations=100, c_lambda=0.05, c_alpha=40)
df = pd.read_csv(resource_filename(testfm.__name__, "data/movielenshead.dat"), sep="::", header=None,
names=["user", "item", "rating", "date", "title"])
training, testing = testfm.split.holdoutByRandom(df, 0.7)
users = {user: list(entries) for user, entries in training.groupby("user")["item"]}
tf.fit(training)
map1 = evaluator.evaluate_model(tf, testing) # map of the original model
#now we try to replace the original factors with on the fly computed factors
#lets iterate over the training data of items and the users
for u, items in users.items():
#user id in the tf
uid = tf.data_map[tf.get_user_column()][u] # user id
iids = [tf.data_map[tf.get_item_column()][i] for i in items] # item ids that user has seen
#original_factors = tf.factors["user"][uid]
new_factors = pyTF.online_user_factors(tf.factors[1], iids, p_param=40, lambda_param=0.05)
#replace original factors with the new factors
tf.factors[0][uid, :] = new_factors
#tf.update_user_factors(uid, new_factors)
#lets evaluate the new model with real-time updated factors
map2 = evaluator.evaluate_model(tf, testing)
#The difference should be smaller than 20%
assert abs(map1[0]-map2[0]) < 0.2*map1[0]
def test_user_model_update(self):
pyTF = PyTensorCoFi()
Y = np.array([[-1.0920831, -0.01566422], [-0.8727925, 0.22307773], [0.8753245, -0.80181429],
[-0.1338534, -0.51448172], [-0.2144651, -0.96081265]])
user_items = [1, 3, 4]
res = pyTF.online_user_factors(Y, user_items, p_param=10, lambda_param=0.01)
assert np.array([-1.18324547, -0.95040477]).all() == res.all(), "Results not equal"
def test_user_model_update(self):
pyTF = PyTensorCoFi()
Y = np.array([[-1.0920831, -0.01566422], [-0.8727925, 0.22307773],
[0.8753245, -0.80181429], [-0.1338534, -0.51448172],
[-0.2144651, -0.96081265]])
user_items = [1, 3, 4]
res = pyTF.online_user_factors(Y,
user_items,
p_param=10,
lambda_param=0.01)
assert np.array([-1.18324547,
-0.95040477]).all() == res.all(), "Results not equal"
def test_dynamic_updates(self):
"""
TensorCoFi dynamic update
We will take a tensor cofi. Train the model, evaluate it. Then we remove all the user factors
and recompute them using the online_user_factors to check if the performance is almost the same...
"""
pyTF = PyTensorCoFi(n_factors=20,
n_iterations=5,
c_lambda=0.05,
c_alpha=40)
evaluator = Evaluator()
tf = TensorCoFi(n_factors=2,
n_iterations=100,
c_lambda=0.05,
c_alpha=40)
df = pd.read_csv(resource_filename(testfm.__name__,
"data/movielenshead.dat"),
sep="::",
header=None,
names=["user", "item", "rating", "date", "title"])
training, testing = testfm.split.holdoutByRandom(df, 0.7)
users = {
user: list(entries)
for user, entries in training.groupby("user")["item"]
}
tf.fit(training)
map1 = evaluator.evaluate_model(tf,
testing) # map of the original model
#now we try to replace the original factors with on the fly computed factors
#lets iterate over the training data of items and the users
for u, items in users.items():
#user id in the tf
uid = tf.data_map[tf.get_user_column()][u] # user id
iids = [tf.data_map[tf.get_item_column()][i]
for i in items] # item ids that user has seen
#original_factors = tf.factors["user"][uid]
new_factors = pyTF.online_user_factors(tf.factors[1],
iids,
p_param=40,
lambda_param=0.05)
#replace original factors with the new factors
tf.factors[0][uid, :] = new_factors
#tf.update_user_factors(uid, new_factors)
#lets evaluate the new model with real-time updated factors
map2 = evaluator.evaluate_model(tf, testing)
#The difference should be smaller than 20%
assert abs(map1[0] - map2[0]) < 0.2 * map1[0]
