def general_CFV(data, prediction_function, target_column, param, k=10):
"""
:param data: data to perform CFV on
:param prediction_function: prediction_function: maps (training data, testing data, k)
to a prediction series of testing data
:param target_column: column of df that is being predicted
:param param: parameter of the model being used
:param k: k to be used in k fold CFV
:return: CFV r squared
"""
train_valid_splits = tools.train_valid_k_fold_sets(data, k)
total_r_sq = 0
# iterate through each (train,valid) tuple finding r squared for each
# add this r squared to the running total r squared
for train_set, valid_set in train_valid_splits:
# find predictions of valid_set based on train_set
valid_set_predictions = prediction_function(train_set, valid_set,
param)
valid_set_r_sq = tools.r_squared(valid_set, target_column,
valid_set_predictions)
total_r_sq += valid_set_r_sq
# find the average r squared over the different validation sets
avg_r_sq = total_r_sq / len(train_valid_splits)
return avg_r_sq
def lasso_CFV(train, features, l2_pen, k):
my_sets = tools.train_valid_k_fold_sets(train, k)
total_r_sq = 0
for t, v in my_sets:
r_model = Lasso(alpha=l2_pen, normalize=True, max_iter=MAX_ITERATIONS)
r_model.fit(t[features], t["price"])
r_sq = r_model.score(v[features], v["price"])
total_r_sq += r_sq
avg_r_sq = total_r_sq / k
return avg_r_sq
def ridge_CFV(train, features, l2_pen, k):
my_sets = tools.train_valid_k_fold_sets(train, k)
total_r_sq = 0
for t, v in my_sets:
r_model = Ridge(alpha=l2_pen, normalize=True)
r_model.fit(t[features], t["price"])
r_sq = r_model.score(v[features], v["price"])
total_r_sq += r_sq
avg_r_sq = total_r_sq / k
return avg_r_sq
def CFV_r_sq(data, lam, k=10):
"""
Does (k = 10) fold cross validation of the data
:param training:
:return:
"""
my_sets = tools.train_valid_k_fold_sets(data, k)
total_r_sq = 0
for t, v in my_sets:
v_predictions = data_predictions(t, v, lam)
v_r_sq = tools.r_squared(v, "price", v_predictions)
total_r_sq += v_r_sq
avg_r_sq = total_r_sq / k
return avg_r_sq
def cross_val_lambda(train, lam, k=10):
"""
Return cross validation r squared
For gaussian kernel regression
:param train:
:param lam:
:param k:
:return:
"""
cross_valid_sets = tools.train_valid_k_fold_sets(train, k)
total_r_sq = 0
for t,v in cross_valid_sets:
v_r_sq = gauss_r_sq(t,v,lam)
total_r_sq += v_r_sq
avg_r_sq= total_r_sq / k
return avg_r_sq
numeric_features.remove("zipcode")
numeric_features.remove("long")
numeric_features.remove("lat")
#numeric_features.remove("yr_built")
features = numeric_features
"""
find the best l2 penalty
done by finding r squared values for different l2 penalties
then select the l2 penalty that gives the highest r squared value
"""
# choose a value of k for k-fold cross validation
k = 50
# set consisting of (test,valid) pairs for k fold validation
my_sets = tools.train_valid_k_fold_sets(train_data, k)
# find the r squared values for different l2 penalties
# then select the l2 penalty that gives the highest r squared value
l2_pens = [10, 40, 100]
best_l2_value = None
best_cross_r_sq = -np.inf
for l2_pen in l2_pens:
cross_r_sq = lasso_CFV(train_data, features, l2_pen, k)
print("For l2 penalty {}, r squared is {}".format(l2_pen,
np.round(cross_r_sq, 5)))
if cross_r_sq > best_cross_r_sq:
best_cross_r_sq = cross_r_sq
best_l2_value = l2_pen
print()
print("The best value for l2 is", best_l2_value)
best_model = Lasso(alpha=0.03, normalize=True, max_iter=MAX_ITERATIONS)