def lgbm_insight_er():
"""Return 5-fold cross validation scores r2, mae, rmse"""
steps = [('scaler', t.MyScaler(dont_scale='for_profit')),
('knn', t.KNNKeepDf())]
pipe = Pipeline(steps)
pipe.fit(X_raw_er)
X = pipe.transform(X_raw_er)
# Run once to get ideal parameters
# params = {
# 'max_bin': [10, 20, 50, 100, 255],
# 'num_leaves': [5, 10, 31, 50],
# 'min_data_in_leaf': [10, 20, 30],
# 'bagging_fraction': [.1, .3, .5, .7, 1]
# }
# lgb_q = LGBMRegressor(objective='quantile')
# gs = RandomizedSearchCV(lgb_q, params,
# scoring=['r2', 'neg_mean_squared_error',
# 'neg_mean_absolute_error'],
# refit='neg_mean_squared_error'
# )
# gs.fit(X, y_er)
lgbm = LGBMRegressor(num_leaves=50,
max_bin=100,
bagging_fraction=0.1,
objective='quantile')
cv_results = cross_validate(
lgbm,
X.to_numpy(),
y_er,
scoring=['r2', 'neg_mean_squared_error', 'neg_mean_absolute_error'],
return_train_score=True)
output = pd.DataFrame(
{
'train_r2': [cv_results['train_r2'].mean()],
'train_rmse': [
np.mean([
np.sqrt(abs(i))
for i in cv_results['train_neg_mean_squared_error']
])
],
'train_mae':
[abs(cv_results['train_neg_mean_absolute_error'].mean())],
'test_r2': [cv_results['test_r2'].mean()],
'test_rmse': [
np.mean([
np.sqrt(abs(i))
for i in cv_results['test_neg_mean_squared_error']
])
],
'test_mae':
[abs(cv_results['test_neg_mean_absolute_error'].mean())]
},
index=['LGBM'])
return output
def r_out(df, x_vars, y_var, csv_label):
def to_r(x, y, csv_label):
"""merges y and X and exports csv to data/processed for use in R, etc.
Use after scaling/imputing"""
return pd.concat([y, x],
axis=1).to_csv('data/processed/' + csv_label + '.csv',
index=False)
X_raw, y = get_train_test(df, x_vars, y_var)
steps = [('scaler', t.MyScaler(dont_scale=['for_profit'])),
('knn', t.KNNKeepDf())]
pipe = Pipeline(steps)
pipe.fit(X_raw)
X = pipe.transform(X_raw)
to_r(X, y, csv_label)
def lr_insight_wr():
"""Return 5-fold cross validation scores r2, mae, rmse"""
steps = [('scaler', t.MyScaler(dont_scale='for_profit')),
('knn', t.KNNKeepDf())]
pipe = Pipeline(steps)
pipe.fit(X_raw)
X = pipe.transform(X_raw)
lr = LinearRegression()
lr.fit(X, y)
cv_results = cross_validate(
lr,
X,
y,
scoring=['r2', 'neg_mean_squared_error', 'neg_mean_absolute_error'],
return_train_score=True)
output = pd.DataFrame(
{
'train_r2': [cv_results['train_r2'].mean()],
'train_rmse': [
np.mean([
np.sqrt(abs(i))
for i in cv_results['train_neg_mean_squared_error']
])
],
'train_mae':
[abs(cv_results['train_neg_mean_absolute_error'].mean())],
'test_r2': [cv_results['test_r2'].mean()],
'test_rmse': [
np.mean([
np.sqrt(abs(i))
for i in cv_results['test_neg_mean_squared_error']
])
],
'test_mae':
[abs(cv_results['test_neg_mean_absolute_error'].mean())]
},
index=['LR'])
return output
def resample_kfold(X, y, sparse_df, model, folds=5, prefix='name_'):
# Get folds
folds = ShuffleSplit(n_splits=folds, test_size=(1 / folds))
folds.get_n_splits(X, y)
# Set up history/scoring lists
model_history = []
train_r2 = []
train_mse = []
train_mae = []
test_r2 = []
test_mse = []
test_mae = []
# execute k-fold
for train_index, test_index in folds.split(X, y):
xtrain, xtest = X.iloc[train_index, :], X.iloc[test_index, :]
ytrain, ytest = y[train_index], y[test_index]
# Simulate missingness on test fold
xtest = pd.DataFrame(xtest, columns=X.columns)
xtest = sim_miss(xtest, sparse_df)
# Scale/transform xtrain
steps = [('scaler', t.MyScaler(defs.dummy_vars)),
('knn', t.KNNKeepDf())]
pipe = Pipeline(steps)
pipe.fit(xtrain)
xtrain = pipe.transform(xtrain)
# scale/impute test (test has simulated missing, imputed on
# data from training folds)
xtest = pipe.transform(xtest)
# Run the model
loop_model = copy.copy(model)
loop_model.fit(xtrain, ytrain)
# Save models
model_history.append(loop_model)
# Save Performance
train_r2.append(r2_score(ytrain, loop_model.predict(xtrain)))
train_mse.append(mean_squared_error(ytrain,
loop_model.predict(xtrain)))
train_mae.append(
mean_absolute_error(ytrain, loop_model.predict(xtrain)))
test_r2.append(r2_score(ytest, loop_model.predict(xtest)))
test_mse.append(mean_squared_error(ytest, loop_model.predict(xtest)))
test_mae.append(mean_absolute_error(ytest, loop_model.predict(xtest)))
performance = {
prefix + 'train_r2': train_r2,
prefix + 'train_mse': train_mse,
prefix + 'train_mae': train_mae,
prefix + 'test_r2': test_r2,
prefix + 'test_mse': test_mse,
prefix + 'test_mae': test_mae
}
return performance
complete_df = pd.read_pickle('data/interim/complete_df.pickle')
X_raw, y = get_train_test(complete_df,
defs.sparse_vars,
'would_recommend',
return_full=True)
X_raw_er, y_er = get_train_test(complete_df,
defs.sparse_vars,
'RATING_EST',
return_full=True)
X_raw.rename(columns=renames, inplace=True)
X_raw_er.rename(columns=renames, inplace=True)
steps = [('scaler', t.MyScaler(dont_scale='for_profit')),
('knn', t.KNNKeepDf())]
pipe = Pipeline(steps)
pipe.fit(X_raw, defs.dummy_vars)
X = pipe.transform(X_raw)
#%%
# RandomSearch best XGBRegressor parameters
params = {
"learning_rate": [0.05, 0.10, 0.15, 0.20, 0.25, 0.30],
"max_depth": [3, 4, 5, 6, 8, 10, 12, 15],
"min_child_weight": [1, 3, 5, 7],
"gamma": [0.0, 0.1, 0.2, 0.5, 1],
"colsample_bytree": [0.3, 0.4, 0.5, 0.7]
}
rs = RandomizedSearchCV(XGBRegressor(), params)
